# Soren AI: full text

> Soren is an AI transformation partner for traditional businesses. Our engineers embed with your team to put AI into production in weeks, then stay to support it.

Every published post on soren-ai.com/blog, in full. The HTML pages at each URL are authoritative. Index: /llms.txt

---

# Can Lawyers Use ChatGPT With Client Documents?

URL: https://soren-ai.com/blog/can-lawyers-use-chatgpt-with-client-documents
Author: Kevin Xie, Founder & CEO, Soren AI
Published: Oct 30, 2026

What ABA Formal Opinion 512 and state bars say about putting client information into ChatGPT, which ChatGPT plans train on your data, and a checklist for firms.

Sometimes, with care. ABA Formal Opinion 512 says lawyers need a client's informed consent before putting information about the representation into a generative AI tool that can learn from it, and boilerplate in an engagement letter isn't enough. Business plans that don't train on your data lower the risk, but the duties of confidentiality, accuracy and supervision still apply.

Key takeaways:
- Under ABA Formal Opinion 512, client information in a self-learning AI tool needs informed client consent, not boilerplate.
- Consumer ChatGPT can train on conversations unless the user turns that off; OpenAI's business plans don't train on business data by default.
- Whatever the plan, lawyers must review AI output as carefully as the task requires, check every citation before filing, and supervise how staff use the tool.

## What does the ABA say about lawyers using ChatGPT?

The ABA allows it, under the same duties that apply to any tool. Its first formal opinion on generative AI, Formal Opinion 512 (July 2024), applies the existing Model Rules rather than creating new ones.[1]

On confidentiality, the opinion is direct. For tools that can learn from what you enter, "a client's informed consent is required prior to inputting information relating to the representation into such a GAI tool." And "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI" is not enough.[1] That sits on top of Model Rule 1.6(c), which requires reasonable efforts to prevent unauthorized disclosure of client information.[2]

Table: What ABA Formal Opinion 512 means for everyday ChatGPT use
| Duty | Model Rule | What it means in practice |
| --- | --- | --- |
| Competence | 1.1 | Understand what the tool can and can't do, and check its work |
| Confidentiality | 1.6 | Know where client information goes; get informed consent for self-learning tools |
| Communication | 1.4 | Explain AI use when the client asks, the engagement terms or outside counsel guidelines require it, it affects the fee, or its output will influence a significant decision |
| Candor and accuracy | 3.1, 3.3, 8.4(c) | Never file AI-drafted work without verifying citations and facts |
| Supervision | 5.1, 5.3 | Set firm policies and train lawyers and staff who use AI |
| Fees | 1.5 | Bill reasonable fees for time actually spent; generally don't bill for learning the tool |

On fees, the ABA's announcement put it plainly: in most circumstances, a lawyer cannot charge a client for learning how to work a generative AI tool.[3]

## What do state bars say about client information in AI tools?

State bars broadly agree with the ABA, and some are stricter. Your state's rules and opinions control, so check them before setting firm policy.

- California: the State Bar's 2026 practical guidance says that, as a general matter, a lawyer "must not input any confidential information of the client into a generative AI solution that may present material risks to confidentiality or security, absent informed client consent" and that reasonable efforts require more than reliance on a vendor's marketing assurances.[4]
- Florida: Ethics Opinion 24-1 (January 2024) recommends getting the affected client's informed consent before using a third-party generative AI program if it would involve disclosing confidential information.[5]
- North Carolina: 2024 Formal Ethics Opinion 1 says that, as of its date, "lawyers should avoid inputting client-specific information into publicly available AI resources."[6]

## Does it matter which version of ChatGPT you use?

Yes. The biggest difference is whether your conversations can be used to train OpenAI's models and who controls how long they are kept.

Table: How ChatGPT plans handle your data, per OpenAI's published policies
|  | Consumer ChatGPT (Free, Plus, Pro) | ChatGPT Business, Enterprise and the API |
| --- | --- | --- |
| Used to train models | Yes, unless the user turns off "Improve the model for everyone" | Not by default |
| Who controls retention | The individual user | Enterprise customers can set retention; eligible API customers can get zero data retention |
| Firm oversight | None: it's a personal account | Admin controls across the firm |
| Fit for client information | Avoid | Possible, with consent where required and a firm policy |

OpenAI's data controls page says consumer ChatGPT may use conversations to improve its models unless the user opts out.[7] Its enterprise privacy page says it does not use business data from ChatGPT Business, Enterprise, Edu or the API for training by default.[8]

Retention can also be taken out of your hands. In 2025, a court order in the New York Times lawsuit required OpenAI to preserve consumer and Team conversations, and API data without zero data retention, that it would otherwise have deleted. The order did not cover ChatGPT Enterprise or Edu. OpenAI says its obligations under it ended on September 26, 2025, though it still holds limited April to September 2025 data the Times is seeking.[9] The lesson for firms: a vendor's deletion policy is not the last word on where client data ends up.

## What about mistakes and made-up citations?

Verification is the lawyer's job, every time. In Mata v. Avianca (S.D.N.Y. 2023), a court sanctioned two lawyers and their firm, imposing a $5,000 penalty, after they submitted non-existent judicial opinions with fake quotes and citations created by ChatGPT and then stood by them.[10]

It keeps happening. An independent database maintained by researcher Damien Charlotin listed more than 2,200 court decisions worldwide involving AI-fabricated content as of late October 2026.[11] A business plan protects confidentiality. It does nothing for accuracy.

## How many lawyers already use AI tools like ChatGPT?

About three in ten, as of 2024. In the ABA's 2024 AI TechReport, 30.2% of attorneys said their offices use AI tools, from 17.7% of solo practitioners to 47.8% at firms of 500 or more lawyers. ChatGPT was the tool respondents most often said they had adopted or were seriously considering.[12]

Without a firm-approved option, some people will use personal accounts. That is the riskiest version of all, because nobody at the firm can see what went in.

## What should a firm do before lawyers use ChatGPT with client documents?

Treat it like any other vendor that touches client information, then add the AI-specific checks.

1. Provide an approved tool. Use a business plan that doesn't train on your data (on OpenAI, retention controls come with Enterprise, not Business), and ban personal accounts for client work.
2. Get informed consent where it's required. Make it specific to the tool and the matter, not a boilerplate engagement-letter clause.
3. Share only what the task needs. Remove names and identifiers when the work doesn't depend on them.
4. Verify carefully. Review output with care that fits the task, and check every citation, quote and fact before anything is filed or sent to a client.
5. Write a short AI policy and train people on it. Cover approved tools, what may go in, who reviews output, and how to report a problem.
6. Bill honestly. Charge reasonable fees for time actually spent, and generally not for learning the tool.
7. Check your state bar. Read your jurisdiction's opinions and recheck them as they change.

This is general information, not legal advice. Your state's rules and your firm's ethics counsel have the final word.

## When does a firm need more than ChatGPT?

When the work has to happen inside your own systems. ChatGPT is a good drafting and research assistant, but it isn't set up to run your firm's workflows, like opening a matter from an intake form or keeping your practice management system up to date.

For firms that want AI working on real matters, Soren AI builds it into the systems they already use, deployed in the cloud, on premises or hybrid, depending on where the firm's policies allow client data to go. Common first projects are client intake and a searchable internal knowledge base. More on AI for law firms (https://soren-ai.com/industries/legal) and how we work (https://soren-ai.com/how-we-work). If data location is the sticking point, the same questions apply as when banks keep customer data in-region (https://soren-ai.com/blog/bank-ai-data-residency).

Use AI on client work without the guesswork: Start with a 30-minute call with our team. We'll look at your systems, your biggest bottlenecks and the quickest wins, starting with how client information moves through your firm.

## Frequently asked questions

### Can I use ChatGPT for legal research?

Yes, as a starting point, but verify everything. ChatGPT can summarize issues and suggest directions, and it can also invent cases and quotes. Courts have sanctioned lawyers for filing AI-fabricated citations. Confirm every authority in a trusted research service before relying on it or citing it.

### Is the free version of ChatGPT safe for client work?

Generally no. Consumer ChatGPT can use conversations to train OpenAI's models unless the user turns that setting off, and the firm has no oversight of a personal account. North Carolina's State Bar, for one, says lawyers should generally avoid putting client-specific information into publicly available AI tools. Use a firm-approved business plan instead.

### Does ChatGPT Enterprise make it okay to upload client documents?

It lowers the risk, because OpenAI doesn't train on Enterprise data by default and the firm controls retention. It doesn't remove your duties. You may still need informed client consent, depending on the tool and your state, and you still have to verify the output and supervise how people use it.

### Do lawyers need client consent to use AI?

Not for every use. ABA Formal Opinion 512 says informed consent is required before putting information about the representation into a generative AI tool that can learn from it, and says boilerplate engagement-letter language isn't enough. For other uses, consent depends on the risk and on your state's rules.

## Sources

1. Formal Opinion 512: Generative Artificial Intelligence Tools, American Bar Association, July 2024: https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
2. Rule 1.6: Confidentiality of Information, ABA Model Rules of Professional Conduct: https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_1_6_confidentiality_of_information/
3. ABA issues first ethics guidance on a lawyer's use of AI tools, American Bar Association, July 2024: https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/
4. Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, State Bar of California, 2026: https://www.calbar.ca.gov/sites/default/files/portals/0/documents/ethics/Generative-AI-Practical-Guidance.pdf
5. Ethics Opinion 24-1, The Florida Bar, January 2024: https://www.floridabar.org/etopinions/opinion-24-1/
6. 2024 Formal Ethics Opinion 1, North Carolina State Bar, November 2024: https://www.ncbar.gov/for-lawyers/ethics-and-governing-rules/ethics-opinions/opinions/2024-formal-ethics-opinion-1/
7. Data controls in ChatGPT, OpenAI, October 2026: https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt
8. Enterprise privacy at OpenAI, OpenAI, January 2026: https://openai.com/enterprise-privacy/
9. How we're responding to The New York Times' data demands, OpenAI, updated October 2025: https://openai.com/index/response-to-nyt-data-demands/
10. Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023), opinion and order on sanctions: https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/
11. AI Hallucination Cases database, Damien Charlotin, updated October 2026: https://www.damiencharlotin.com/hallucinations/
12. 2024 Artificial Intelligence TechReport, American Bar Association: https://www.americanbar.org/groups/law_practice/resources/tech-report/2024/2024-artificial-intelligence-techreport/

---

# What Is an AI Agent? A Plain-English Guide for Operators

URL: https://soren-ai.com/blog/what-is-an-ai-agent
Author: Kevin Xie, Founder & CEO, Soren AI
Published: Oct 29, 2026

An AI agent is software that works toward a goal, picks its next step and uses your systems to get there. What agents can and can't do, and where to start.

An AI agent is software that uses an AI model to work toward a goal, deciding its next step as it goes and using tools such as your document system, email or case management software. A chatbot answers questions. An agent can read an intake form, check a record, draft a reply and route it to a person to approve.

Key takeaways:
- An AI agent decides its own next step toward a goal; traditional automation follows a fixed script.
- Agents fit messy, judgment-heavy back-office work like intake and document review, with people approving what matters.
- Many tasks don't need an agent at all. A simpler, predictable workflow is often the better choice.

## What is an AI agent, in plain English?

An AI agent is software that is given a goal and figures out the steps to reach it. Google Cloud defines agents as software systems that use AI to pursue goals and complete tasks on behalf of users.[1] The important part is the second half: an agent takes actions as well as answering.

Three parts make something an agent:

1. A model that reasons. A large language model reads the situation and decides what to do next.
2. Tools. Access to the systems where work happens: reading documents, searching a database, filling in a form, sending an email.
3. A goal and limits. A clear task, such as "prepare this new matter for conflicts review," plus rules about what it may do on its own and what needs a person.

Anthropic draws a useful line here. In a workflow, the AI model is used inside steps a developer has fixed in advance. In an agent, the model directs its own process and decides which tools to use.[2] Many real systems combine the two.

## How is an AI agent different from a chatbot or traditional automation?

A chatbot answers, traditional automation repeats, and an agent works toward an outcome. The differences matter because vendors use the word loosely.

Table: Chatbots, traditional automation and AI agents compared
|  | Chatbot or assistant | Traditional automation (RPA) | AI agent |
| --- | --- | --- | --- |
| What it does | Answers questions and drafts text when asked | Repeats the same clicks and keystrokes | Works through a multi-step task toward a goal |
| How it decides | Responds to each prompt | Follows fixed rules written in advance | Chooses its next step based on what it finds |
| Messy inputs | Handles them, but only inside the chat | Can break when a form or screen changes | Handles varied documents and exceptions |
| Uses your systems | Rarely | Yes, through the screen or an integration | Yes, through tools it is allowed to use |
| Back-office example | Summarizes an engagement letter you paste in | Copies fields from one system to another | Reads an intake email, checks the client list and drafts the new matter for review |

Be skeptical of labels. Gartner calls the rebranding of existing assistants, robotic process automation and chatbots without substantial agentic capabilities "agent washing," and estimated in 2025 that only about 130 of the thousands of vendors claiming agentic AI were offering the real thing.[3]

## What can an AI agent actually do in a back office?

The best fits are repetitive tasks that still need some reading and judgment: work that is too varied for a rigid script but too routine for your most experienced people.

- Client intake at a law firm: read the inquiry, pull out the parties, check them against existing clients, open a draft matter and send it to an attorney for approval. See more on AI for law firms (https://soren-ai.com/industries/legal).
- Loan document review at a bank: check an application package for missing pages, extract the figures and flag anything that doesn't match.
- Prior authorization prep in healthcare: gather the records a payer asks for and assemble the request for staff to review and submit.
- A searchable knowledge base: answer staff questions from the firm's own policies, templates and past work, with links to the source documents.

Client intake and an internal knowledge base are two of the most common first workflows we build at Soren AI. At one law firm, intake that used to take three days now takes about 20 minutes.

## When don't you need an AI agent?

When the task follows the same steps every time. If the inputs are clean and the rules never change, a simple, predictable workflow is cheaper, faster and easier to audit.

The AI labs say the same. OpenAI's guide to building agents notes that for some tasks a deterministic solution may suffice,[4] and Anthropic recommends starting with the simplest approach that works, because agentic systems are often slower and cost more in exchange for better task performance.[2]

A good test: if you could write the process down as a checklist with no "it depends," you probably don't need an agent.

## How many companies are actually using AI agents?

Fewer than the hype suggests, and mostly large ones. In McKinsey's 2026 State of AI survey, 40% of respondents at organizations with more than $1 billion in revenue reported scaling AI agents, up from 27% a year earlier. Among respondents at smaller organizations, the share held flat at 22%.[5]

In August 2025, Gartner forecast that up to 40% of enterprise applications would include task-specific agents by 2026, up from less than 5% at the time.[6] It also predicted that more than 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or weak risk controls, and advised pursuing agentic AI only where it delivers clear value.[3]

## What are the risks of AI agents?

The main risks are agents acting beyond their authority, being manipulated by what they read, and mistakes nobody catches. FINRA's 2026 oversight report warns that agents may act beyond the user's actual or intended scope and authority, and suggests firms consider where to put human-in-the-loop oversight.[7]

Security is a real concern too. In NIST testing published in January 2025, the strongest new attack that hides malicious instructions in data an agent reads succeeded 81% of the time, versus 11% for the strongest baseline attack.[8]

The controls that help are practical:

- Least access: give the agent only the systems and permissions the task needs, starting read-only where possible.
- People approve what matters: anything sent to a client, filed, paid or recorded goes to a person first.
- A full log: keep a record of what the agent read, decided and did, so you can review and explain it.
- Know where data goes: map where prompts, documents and logs are processed and stored. For banks, we covered how to keep customer data in-region (https://soren-ai.com/blog/bank-ai-data-residency); for law firms, whether lawyers can use ChatGPT with client documents (https://soren-ai.com/blog/can-lawyers-use-chatgpt-with-client-documents).

## How should an operations team start with AI agents?

Start with one workflow. Pick a task that is frequent, well understood and easy to measure.

1. Map the workflow as it really runs, including the exceptions and the systems it touches.
2. Decide what the agent may do alone and what always goes to a person.
3. Start read-only or draft-only, so the agent prepares work and people send it.
4. Measure one thing, such as turnaround time or hours saved, against how the work ran before.
5. Expand only after it holds up on real work for a few weeks.

At Soren, engineers spend the first two weeks interviewing the people who do the work, and the first workflow is typically live within two to three weeks of getting access to the systems it needs. More on how we work (https://soren-ai.com/how-we-work), and on how this differs from traditional consulting (https://soren-ai.com/blog/ai-native-vs-traditional-consulting).

Find the right first workflow for an agent: Start with a 30-minute call with our team. We'll look at your back office and tell you honestly whether an agent, a simpler workflow or nothing at all is the right fit.

## Frequently asked questions

### Is ChatGPT an AI agent?

A standard ChatGPT conversation is an assistant: it answers and drafts when you ask. Some AI products now add agent features that can browse, use tools and take actions. Whether something is an agent depends on whether it works through steps toward a goal on its own, not on the brand name.

### Will AI agents replace back-office staff?

In practice, agents take over the repetitive parts of a role, like retyping, checking for missing documents and chasing information. People keep the judgment calls, the client relationships and the final approval. The goal is usually to clear backlogs and handle more work, not to replace the people who know the process.

### Are AI agents safe for regulated industries?

They can be, with the right controls: limited system access, human approval for anything consequential, full logs, and a clear map of where data is processed and stored. FINRA's 2026 oversight report, for example, suggests broker-dealers consider where human-in-the-loop oversight belongs. Build that oversight in from the start.

### How long does it take to deploy an AI agent?

A focused agent for one workflow can be live in a few weeks if the systems and data it needs are accessible. At Soren, the first workflow is typically live within two to three weeks of getting access. Broader rollouts take longer because each new task needs its own testing.

## Sources

1. What are AI agents?, Google Cloud, 2026: https://cloud.google.com/discover/what-are-ai-agents
2. Building effective agents, Anthropic, December 2024: https://www.anthropic.com/engineering/building-effective-agents
3. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner, June 2025: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
4. A practical guide to building agents, OpenAI, 2025: https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
5. The state of AI in 2026: On the road to ROI, McKinsey & Company, August 2026: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
6. Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Gartner, August 2025: https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025
7. 2026 Annual Regulatory Oversight Report, FINRA, December 2025: https://www.finra.org/sites/default/files/2025-12/2026-annual-regulatory-oversight-report.pdf
8. Technical Blog: Strengthening AI Agent Hijacking Evaluations, NIST, January 2025: https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations

---

# How Banks Can Use AI and Keep Customer Data In-Region

URL: https://soren-ai.com/blog/bank-ai-data-residency
Author: Kevin Xie, Founder & CEO, Soren AI
Published: Oct 28, 2026

Yes, banks can use AI and keep customer data in-region. What in-region has to cover, the three deployment options, and what examiners will ask to see.

Yes. A bank can use AI and keep customer data in-region, but in-region has to cover the whole workflow, not just stored files. Check where prompts are processed, where answers and logs land, who can access them, and what happens during failover. Choose a deployment only after that data path is clear.

Key takeaways:
- Storage location alone doesn't settle data residency: prompts, processing, logs, backups, failover and support access all count.
- A vendor's regional service, the bank's own cloud and on-premises infrastructure each change who controls the data path and who runs it.
- Pilot in stages: internal policies first, then one approved, read-only customer task, with staff reviewing every output.

## Why does data residency matter for banks using AI?

Because the bank stays accountable for customer information even when an AI vendor or cloud provider handles part of the work. A regional storage setting is only one control. The bank has to know where data is processed, who can reach it, and which party owns each safeguard, a responsibility the FFIEC's statement on cloud computing makes clear.[1] It is also what bankers worry about most: in a 2026 survey of 17 community banks, 55% of respondents named security and data leakage as their top AI challenge.[2]

Picture a banker summarizing a customer file. The file may stay in the bank's cloud account while an excerpt travels to a model, an answer comes back, and a log is saved. Each step can have its own location and retention settings. Microsoft's documentation, for example, keeps stored data in the chosen geography, while prompts sent to a "Global" deployment may be processed in any Azure region.[3] "Stored in-region" does not prove the whole workflow stays there.

So define the boundary first: one cloud region, one country, or a wider geography. Then trace source records, prompts, outputs, logs, backups, failover, support access and subcontractors. If one step crosses the boundary, a regional label can't settle the question. Recheck the map whenever a provider changes its model, routing, logging or failover setup. Law firms face the same question with client files; we covered whether lawyers can use ChatGPT with client documents (https://soren-ai.com/blog/can-lawyers-use-chatgpt-with-client-documents).

## What are the main deployment options?

There are three practical routes: a vendor-operated service with a regional commitment, an AI application in the bank's own cloud environment, or infrastructure on premises. The real question is who controls each part of the data path and who maintains it.

Table: Three ways to deploy AI with an in-region data boundary
|  | Vendor-operated, regional commitment | In the bank's cloud | On premises |
| --- | --- | --- | --- |
| What the bank controls | Vendor choice, contract terms, user access and available settings; the vendor runs the service | More of the application, identity, configuration and data settings; the cloud provider runs the infrastructure | Hardware, network and most operational settings, if the full workflow really runs there |
| What to check | Does the commitment cover processing, logs, backups, failover, support access and subcontractors? | Do managed model calls or connected services cross the boundary? Who owns each security control? | Are there external model calls or remote support paths? Can the bank keep up security, capacity and recovery? |
| Cost and speed | Fastest to start and no hardware; usage fees, vendor review and contract oversight remain | Moderate setup; more flexibility, plus engineering and ongoing cloud operations for the bank | Slowest to start; the bank buys, houses, staffs and scales the hardware |

"Private cloud" doesn't necessarily mean on premises. NIST's definition allows a private cloud to sit on or off premises and to be run by the organization, a third party, or both.[4] Ask where each workload actually runs and who owns the controls; the label isn't an answer.

## What should a bank be ready to show examiners and auditors?

A decision trail: what the AI does, how customer data moves, who owns the controls, and how those controls are tested over time. Interagency guidance on third-party relationships calls for oversight scaled to the activity and its risk, including "ongoing monitoring throughout the duration of a third-party relationship."[5] This is a practical evidence list, not an examiner script.

1. Purpose and risk: the task, the customer data it uses, and what happens if the AI is wrong or unavailable.
2. Data path: where records, prompts, outputs, logs and backups are stored and processed, including failover, subcontractors and any geographic terms in the contract.
3. Access and ownership: who can see data or change settings, including provider staff, and which controls the bank operates itself.
4. Evidence: due diligence, relevant audit reports, tests and monitoring records. Confirm each audit report covers the service and controls the bank relies on.
5. Incidents and exit: how the bank detects problems, restores service, and retrieves or deletes its data if it changes providers.

A data-flow diagram, the vendor's terms and day-to-day operating evidence should all support the same conclusion.

## Where should a bank start with an AI pilot?

Start where no customer data is at risk. Point the AI at the bank's internal policies first. Once you've confirmed data stays in-region, move to one approved task where the AI reads customer documents but can't change them, with staff reviewing every output.

Table: A two-stage pilot
|  | Stage 1: policy search | Stage 2: one customer-document task |
| --- | --- | --- |
| Data used | Internal policies and procedures | Approved customer files, read-only |
| What the AI does | Answers questions, like the callback procedure for an outgoing wire, citing the exact policy section | Drafts a summary of a borrower's financial statements for a credit analyst, or flags gaps |
| Who checks it | The person asking, against the cited policy | The staff owner of the task, every time |
| Move on when | The data path, logs and access controls match the map | Reviewers' corrections are logged and the error pattern is understood |

Why stage it? The revised model risk guidance from the Federal Reserve, OCC and FDIC, SR 26-2 (April 17, 2026, replacing SR 11-7), leaves generative and agentic AI (https://soren-ai.com/blog/what-is-an-ai-agent) out of scope and says a bank's own risk management and governance should set the controls for those tools.[6] For the models it does cover, it warns that "using a model beyond its intended purpose introduces additional uncertainty and risk." That principle is worth borrowing: each expansion, whether a new document type, a new team or write access, gets its own review.

## How does a bank scale AI without losing track of customer data?

Write the data boundary down once, then add one workflow at a time inside it. Score each workflow on two things: how accurate its outputs are, and whether the controls held. Expand only after the people who approved the pilot have reviewed that evidence.

Put the boundary on one page. Which regions data may sit in. Which systems the AI can read from, and which it can never write to. Which kinds of data are allowed. Who can see prompts and logs, and how long logs are kept. Anything a new idea needs that the page doesn't allow goes back to risk and compliance before anyone builds it.

Add workflows in order of data sensitivity. A sensible order for most banks:

1. Internal knowledge: policy and procedure search, with no customer data.
2. Read-only customer documents: checking files for missing items or summarizing them for a reviewer.
3. Drafting for staff: a first draft of a credit memo or exception report that an analyst edits and signs.
4. Anything that writes to a system of record: last, and always with a person approving each change.

Keep two scorecards from day one. A workflow can be accurate and still fail review because nobody can prove where its data went. Agree on the pass mark with your risk team before the pilot starts, not after you've seen the results.

Table: What to track for every AI workflow
|  | What to measure | Evidence to keep | Who reviews it |
| --- | --- | --- | --- |
| Accuracy | How often reviewers change or reject outputs, and why | A sample of outputs checked against the reviewer's final version | The business owner, such as the credit or operations lead |
| Controls | Data stayed in-region, access matched permissions, no writes outside scope | Logs, access reviews, incident records | Information security and compliance |
| Ongoing | Accuracy and controls still hold months later | Quarterly re-checks on a fresh sample | Risk management and internal audit |

Expand only after the review signs off. The gate to the next workflow is a meeting with a packet: the boundary page, both scorecards, and anything that went wrong and how it was fixed. If the packet is thin, the answer is not yet. The same interagency guidance is clear that using a third party "does not diminish" the bank's own responsibility.[5] For a shared vocabulary, NIST's voluntary AI Risk Management Framework, released in January 2023 and now being revised, groups this work into four functions: govern, map, measure and manage.[7]

## How does Soren help a bank scale this way?

Soren AI's engineers embed with your team, so the people who built the pilot are the ones who put together the review packet and answer your risk team's questions. We build into the systems you already run, like Fiserv or Jack Henry cores and nCino, instead of moving your data somewhere new, which keeps the boundary page short. We've shipped cloud, on-premises and hybrid deployments.

We also stay after launch, so the quarterly re-checks aren't handed to someone who has never seen the workflow. Your staff approve every output and make every decision; we make sure there's evidence that they did. More on the model in how we work (https://soren-ai.com/how-we-work), and on our work with banks and credit unions (https://soren-ai.com/industries/finance).

Map your data boundary with us: Start with a 30-minute call with our team. We'll walk through one workflow and where its data would need to stay.

## Frequently asked questions

### Can a community bank use ChatGPT at all?

Yes, for work that doesn't involve customer data, under an enterprise agreement with retention and training settings configured and reviewed by risk and compliance. Putting customer information into a consumer chatbot is different: no bank contract covers it, and the bank couldn't show an examiner where that data went.

### What does "in-region" actually guarantee?

Often only where data is stored. Processing, logs, backups, failover and support access can each sit somewhere else unless the contract and settings say otherwise. Ask the vendor about each step separately, get the answers in writing, and recheck them whenever the provider changes its models or routing.

### Do we need our own GPUs to keep data in-region?

No. Only an on-premises deployment needs the bank's own hardware. A vendor service with a regional commitment, or a model running in the bank's own cloud account, can keep data in-region without the bank buying or running GPUs, as long as every step of the data path stays inside the boundary.

### Does SR 26-2 apply to generative AI?

Not directly. The April 2026 model risk guidance from the Federal Reserve, OCC and FDIC leaves generative and agentic AI out of its scope. It says a bank's own risk management and governance should set the controls for those tools, so banks still need a documented process for approving and monitoring them.

## Sources

1. Joint Statement on Risk Management for Cloud Computing Services, FFIEC, April 2020: https://www.fdic.gov/news/financial-institution-letters/2020/fil20052a.pdf
2. Survey Results: 2026 AI Adoption & Maturity in Banking, Wolf & Company, May 2026: https://www.wolfandco.com/?p=13114
3. Azure OpenAI deployment types, Microsoft Learn, updated August 2026: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/deployment-types
4. SP 800-145: The NIST Definition of Cloud Computing, NIST, 2011: https://csrc.nist.gov/pubs/sp/800/145/final
5. Interagency Guidance on Third-Party Relationships: Risk Management (SR 23-4), June 2023: https://www.federalreserve.gov/supervisionreg/srletters/sr2304.htm
6. SR 26-2: Revised Guidance on Model Risk Management, Federal Reserve, OCC and FDIC, April 2026: https://www.federalreserve.gov/supervisionreg/srletters/sr2602.htm
7. AI Risk Management Framework, NIST: https://www.nist.gov/itl/ai-risk-management-framework

---

# AI-Native vs. Traditional Consulting: What's Different?

URL: https://soren-ai.com/blog/ai-native-vs-traditional-consulting
Author: Kevin Xie, Founder & CEO, Soren AI
Published: Oct 28, 2026

An AI-native consultancy ships working software into your workflows in weeks and stays to run it. A traditional one delivers a plan. When to choose each.

An AI-native consultancy is a consulting firm whose consultants are engineers: they build working AI software into your workflows in weeks, then stay to run and improve it. A traditional consultancy delivers analysis, a strategy and a roadmap, and someone else builds it. Choose by what you need at the end: a plan, or a working system.

Key takeaways:
- An AI-native consultancy delivers working software in production. A traditional consultancy delivers analysis and a plan.
- AI-native firms fit defined back-office workflows such as intake and document review. Traditional firms fit organization-wide change and regulatory strategy.
- Before hiring either, ask who writes the code, who runs it after launch, and what your team owns at the end.

## What is an AI-native consultancy?

In an AI-native consultancy, the consultants are also the engineers. They work alongside your team, build inside your systems, and stay until the software does what it was meant to do. This is the forward-deployed engineer model, which started at Palantir and has spread quickly. McKinsey now publishes an explainer on the role,[1] and job postings for forward-deployed engineers rose more than 800% between January and September 2025, according to Indeed data.[2]

The attention is practical. AI is easy to demo and hard to put into daily use, and embedded engineers exist to close that gap. The Pragmatic Engineer describes the same pull from the engineering side: the work sits between the product and the customer's real systems.[3]

## How does an AI-native consultancy compare to a traditional consulting firm?

The biggest difference is what you have at the end: a plan, or a working system.

Table: Traditional and AI-native consulting, side by side
|  | Traditional consultancy | AI-native consultancy |
| --- | --- | --- |
| Team | Strategists and analysts led by a partner | Engineers embedded directly with your staff |
| First result | A plan; building starts afterward, often with another vendor | A first workflow live in weeks |
| Approach | Assess, recommend, then plan the rollout | Build the first workflow, learn from real use, then expand |
| Deliverables | Analysis, strategy, roadmap, business case | Working software in your workflow, plus monitoring and upkeep |
| After launch | Handoff to an implementer or your IT team | The same engineers support and improve it |
| Success measured by | Quality of the plan | Adoption and impact on the workflow |

## What projects is an AI-native consultancy best for?

An AI-native consultancy is best at turning an existing manual workflow into an automated one. For Soren's clients at law firms (https://soren-ai.com/industries/legal), financial firms (https://soren-ai.com/industries/finance) and nonprofits (https://soren-ai.com/industries/nonprofits), that is usually back-office work such as client intake or a searchable internal knowledge base. These projects tend to share three traits:

- A defined workflow: clear inputs and outputs, so everyone can tell when the project is done.
- Existing systems: the software has to work inside them, which is where embedded engineers add the most value.
- Ongoing operation: AI workflows need monitoring and upkeep after launch.

## How would each approach the same problem?

Take a hypothetical credit union whose staff handle loan applications by hand. Applicants send pay stubs, tax returns and identification. Staff check each package for completeness, key the data into the loan system, and chase anything missing. The process is slow, repetitive, and depends on a few experienced people.

Table: One loan intake problem, two engagements
|  | Traditional consultancy | AI-native consultancy |
| --- | --- | --- |
| First phase | Interviews staff and maps the intake process end to end, then assesses where AI could help across lending | Embeds with the lending team, watches real applications move through, and picks intake document review as the first workflow |
| What you receive | A findings report, a prioritized list of AI opportunities, and a business case for intake | A working tool in your environment that reads incoming documents, checks them for completeness, and prepares the data for your loan system |
| Who builds it | A separate vendor or your IT team, after the engagement ends | The same engineers who scoped it, working inside your systems |
| What comes next | Vendor selection, procurement, and implementation planning | Monitoring the tool on live applications, supporting your staff, and scoping the next workflow |

## What does an AI-native engagement look like at Soren?

At Soren, you get the whole engineering team, typically 10 or more engineers, not one person assigned to your account. The first two weeks set the pattern for everything after.

1. Weeks one and two: we meet your team on site or virtually, interview the people doing the work, and learn how it actually runs. By the end of week two there is a clear scope and project goal, and we are already building in the background.
2. The first workflow: we pick low-hanging fruit with a return you can attribute directly, so success is clear rather than arguable.
3. After launch: the same engineers stay to monitor, support and improve what they built. Engagements usually run a year or more, and you own everything we build.

The most common surprise in our assessments is how scattered the information is. Before any automation, most teams need their documents and data pulled into one place. More on the model in how we work (https://soren-ai.com/how-we-work).

## When should you hire a traditional consulting firm instead?

Sometimes a traditional firm is the better call. Pick one when:

- The change spans the whole organization. Operating model redesign, restructuring and change management are people and strategy problems before they are software problems.
- The question is regulatory. If you need a defensible view of how a rule applies to your institution, regulatory specialists are the right resource.
- You need board-level strategy. A multi-year roadmap across many business units benefits from broad, cross-industry experience.

See what this looks like in your back office: Start with a 30-minute call with our team. We'll look at one of your workflows and tell you honestly whether building it is the right next step.

## Frequently asked questions

### Should we hire an AI-native firm or build an in-house team?

An in-house team gives you long-term ownership and deep knowledge of your business, but it means competing for engineers who are scarce: Indeed data showed forward-deployed engineer postings up more than 800% in 2025. A firm can get workflows live while you hire, and a good one leaves your team owning what was built.

### Is an AI-native consultancy the same as a development shop?

Not quite. A development shop builds to a specification you provide, then moves on. An AI-native firm helps decide which workflow to automate first, builds it inside your existing systems, and stays accountable for how it performs after launch, including the monitoring and fixes AI workflows need.

### What about a traditional firm with an AI practice?

Many large firms now have engineers, and some build real systems. Ask the same questions you would ask anyone: who writes the code, who is on call when it breaks, and whether the engagement starts with building or with a strategy phase.

### How do you choose an AI consulting firm?

Ask four questions. Who writes the code, and who is on call when it breaks? What does your team own at the end? Whose platform is it built on, and what would leaving cost? How many clients does each engineer cover at once? A firm that answers all four plainly is building for you, not selling you a plan.

### How long until the first workflow is live?

With an AI-native firm, weeks rather than quarters. At Soren, the first workflow is typically live within two to three weeks of getting access to the systems and information it needs. A traditional engagement usually delivers its plan first, and building starts after that, often with another vendor.

## Sources

1. What is a forward-deployed engineer?, McKinsey & Company, October 2026: https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-a-forward-deployed-engineer
2. Why Forward-Deployed Engineers Are in High Demand, eWeek, November 2025 (Indeed data): https://www.eweek.com/news/openai-anthropic-cohere-ai-hiring/
3. What are Forward Deployed Engineers, and why are they so in demand?, The Pragmatic Engineer: https://newsletter.pragmaticengineer.com/p/forward-deployed-engineers