In-House vs Outsourced AI Development: Which Fits Your Firm

Build an in-house AI team when AI is part of what you sell and there is enough continuous work to keep engineers busy for years. Outsource when you need a few production workflows built and have nobody inside who could build them. For most accounting firms, law firms and mid-size companies, though, where the developers sit matters less than who owns what gets built and who supervises it once it runs.

What is the difference between in-house and outsourced AI development?

In-house AI development means your own employees design, build and run the system, so the code and know-how stay in the firm. Outsourced AI development means an outside company builds it under contract, and what you keep depends on that contract. With agents, the test cases, review step, runbook and cloud account matter as much as code.

Unlike a website, an agent depends on a model that gets retired. On 30 September 2026, Anthropic notified developers that Claude Sonnet 4.5 leaves the Claude API on 30 November 2026; its deprecation policy promises at least 60 days' notice, after which requests to a retired model fail. If the builders have moved on, that migration falls to you.

Whether to build at all or buy a product is a separate decision, covered in build vs buy AI agents.

The four ways to get an AI agent built

Hire a team. AI engineers join your payroll. You own what they make and carry recruiting, key-person risk and the running of the system.
Outsource to a dev shop. An agency or contract developers build to a specification. You own whatever the statement of work says.
Use a platform. A vendor's agent builder or packaged product: the fastest route to something working, rented for as long as you pay.
Work with an implementation partner. An outside team builds inside your systems, alongside people you name, then hands over the code, prompts, test cases and runbook. The working model is forward-deployed engineering. Gaper is in this category, so weigh what follows accordingly.

Side by side: speed, client data, cost, ownership and supervision

	In-house team	Dev shop	Platform	Implementation partner
Speed to first workflow	Slowest, recruiting comes first	Fast start	Fastest to a demo	Fast start, paced by your reviewers
Where client data goes	Your systems	Wherever the shop builds and tests	The vendor's environment	Your systems, if built in your tenant
Cost shape	Salaries and benefits every year	Project fee, then maintenance fees	Seat or usage fees, indefinitely	Project fee, then your own run cost
Who owns the code	You	Whoever the contract assigns it to	The vendor	You, by written assignment
Who supervises outputs	You	Often nobody named	You, inside the vendor's tools	You, with the review step built in
What survives if it ends	Everything, unless people leave	Whatever was handed over	Nothing you can run	Everything that was handed over

Each column fails differently: the engineer resigns, the contract ends and nobody can change a prompt, the platform's price moves, or the handover is a ceremony, not a transfer.

How much does an in-house AI team cost compared with outsourcing?

An in-house team costs salaries and benefits every year whether the work is steady or not, while outsourcing costs a build fee plus the cost of running the system afterwards. For scale, the US Bureau of Labor Statistics puts the median annual wage for software developers at $135,980 in May 2025, before benefits or any AI premium.

One engineer is also not a team. Set that against the build and run cost of the workflows you need, broken down in our guide to AI agent development cost and, for accounting firms, in our implementation cost bands, which include the reviewer's time most estimates leave out.

When does building an in-house AI team make sense?

Build in-house when AI is part of the product you sell or the core of how you compete, when the work is continuous rather than a few projects, and when you can hire and keep someone senior enough to own architecture and evaluation. If all three hold, outside help is a bridge to your own team, not a substitute for it.

Most professional services firms fail the first test: an accounting firm sells judgment, not software. Hiring engineers for its first two agents buys a permanent cost for a temporary need.

When does outsourcing AI development make sense?

Outsource when you need a defined set of workflows built to production, you have no engineering function to manage AI hires, and you can name who inside will own each result afterwards. Firms skip that last condition, and it decides whether outsourcing works.

The evidence leans toward outside help, with caveats. In MIT NANDA's The GenAI Divide: State of AI in Business 2025 (July 2025), external partnerships with learning-capable, customized tools reached deployment about 67 percent of the time, against about 33 percent for internally built tools. The report's partnership category means procuring external tools and co-developing with vendors, not outsourcing a custom build to a specification, so read it as evidence against going it alone rather than as a verdict on dev shops. The sample was 52 organizations, outcomes were self-reported, and the authors caution that the gap may reflect the organizations rather than the approach. The hybrid model, an internal team co-developing with a vendor, had too little data to quantify.

Who owns the code when you outsource AI development?

When you outsource AI development, the code belongs to whoever the contract assigns it to, and if the contract says nothing useful, probably not to you. Under US copyright law a transfer of ownership is valid only if it is in writing and signed by the owner of the rights conveyed, so paying for software does not by itself make it yours.

The "work made for hire" clause many agreements lean on is narrower than it looks: the Copyright Office's Circular 30 explains that a commissioned work is made for hire only if it falls within nine listed categories and the parties sign a written agreement saying so. Custom software does not fit neatly into those, so also ask for an express written assignment.

Apply the Rent-vs-Own test: if the team that built it disappeared tomorrow, could you run it, change it and explain it to a reviewer? You own it only if you hold all of these:

Code and prompts, in a repository your firm controls, covered by a signed assignment.
The runtime: the cloud account and model API keys are in your firm's name.
The evaluation set: test cases proving the agent still works after a change or model migration.
The runbook: what to do when it fails or a model is retired.
The audit trail: what the agent did and who approved it, stored in your systems.

A vendor keeping its pre-existing libraries is fine if your license to them survives. If anything else above stays with the vendor, you are renting; price it as rent.

Checklist of the Rent-vs-Own test: you own an AI agent only if you hold all five of code and prompts under a signed assignment, the runtime in your firm's name, the evaluation set, the runbook and the audit trail. If any of the five stays with the vendor, you are renting and should price it as rent.

What is the hybrid AI development model?

In the hybrid model an outside team builds the first production workflows inside your systems, while named people on your side own and supervise them from day one and take over the run at the end. The partner brings build experience; your firm brings judgment about what a correct output looks like. When you reach the shortlist, the questions to ask an AI implementation partner separate production experience from demos.

We run engagements this way, through the Gaper method: Assess, Scope, Build, Supervise, Hand over. Supervision is designed with the Gaper Ownership Map, which sorts each workflow step into one of three tiers:

Tier	What it means	Example in an accounting firm
Automated	The agent acts; the log is reviewed by exception	Matching cleared bank items
Agent-drafted, human-approved	The agent prepares; a named person approves before anything leaves	Coding invoices, drafting a client query
Human-owned	People decide; the agent may assist but never acts	Tax positions, anything that gets signed

Reference build: illustrative, not a client engagement.

The middle tier holds most of the value, and it fails quietly when nobody is named as the approver. Our guide to human-in-the-loop AI covers how approval gates are designed.

Roles your team keeps when a partner builds

None of these roles needs an AI engineer, only people who know the work.

Executive owner. A partner or COO who sets scope and accepts go-live.
Workflow owner. The person doing the work today, who defines a correct output and owns the evaluation set.
Reviewer. The licensed person who approves the middle tier; budget for their time.
Technical owner. Your IT or security lead, who holds the cloud account and keys and oversees the vendor.

When volume outgrows these four, hire your first AI engineer, who inherits a working system rather than a blank page. If you want help placing your workflows on the map, that is what our free AI assessment is for.

Can an accounting firm let an outside AI team work with client data?

Yes, but each rule that governs client data reaches the outside team, so the contracts, notices and any consents have to be in place before anyone opens a client file. Three regimes apply at once: Section 7216 for tax return information, the AICPA Code for confidential client information, and the FTC Safeguards Rule for service providers.

This is information, not legal advice; have counsel review the documents before you act.

Section 7216. Under 26 CFR 301.7216-3(a), a preparer may not disclose or use a taxpayer's tax return information before obtaining the taxpayer's written consent, unless section 7216 or 301.7216-2 specifically authorizes it. Rev. Proc. 2013-14 sets the format and content of those consents for Form 1040 series filers. And 301.7216-1(b)(2)(i)(B) counts a person who "develops software that is used to prepare or file a tax return" as a tax return preparer, so a team building that kind of tool is inside the regime.

There is also a route that needs no consent. 301.7216-2(d)(2) lets a preparer disclose tax return information to a contractor programming, maintaining, repairing, testing or procuring equipment or software used for tax return preparation, only to the extent the contracted services need it, and only if every individual who receives it gets written notice of sections 6713 and 7216 and their requirements and penalties. Anything beyond that needs consent under 3(a), unless another exception in 301.7216-2 applies.

Location changes the answer. Under 301.7216-2(d)(1), a preparer may disclose without consent to another preparer for auxiliary services that are not substantive determinations or advice affecting tax liability, but only to one located in the United States. The text of (d)(2) does not address location, so confirm the route with counsel before an offshore contractor receives anything. Separately, 301.7216-3(b)(4) generally bars a US preparer from obtaining consent to disclose a Form 1040 series filer's social security number to a preparer outside the United States. The only exception requires an adequate data protection safeguard as defined in IRS guidance, verified in the consent request; otherwise the SSN must be redacted or masked first. Our post on Section 7216 and AI tools covers the consent itself.

The AICPA Code. Interpretation 1.700.040 of the Code of Professional Conduct says that before disclosing confidential client information to a third-party service provider, a member should either contract with the provider to maintain confidentiality, with reasonable assurance its procedures prevent unauthorized release, or obtain the client's specific consent. Interpretation 1.150.040 adds that the client should first be told, preferably in writing, that a provider may be used, and 1.300.040 makes the member responsible for supervising the provider's work.

The Safeguards Rule. The IRS noted in IR-2026-92 (18 August 2026) that "tax and accounting professionals are considered financial institutions" under the Gramm-Leach-Bliley Act. 16 CFR 314.4(f) then requires you to take reasonable steps to select providers capable of maintaining appropriate safeguards, require those safeguards by contract, and periodically assess them.

The practical answer: build inside your own environment, on redacted or synthetic data during development.

Does outsourcing AI development put privilege at risk for a law firm?

It can, if the outside team or its tools receive client material on terms that do not keep it confidential, so the provider must be supervised like any other nonlawyer assistance. ABA Formal Opinion 512, issued 29 July 2024, applies lawyers' existing duties, including confidentiality and supervision, to generative AI tools and providers.

Opinion 512 says managerial lawyers must establish clear policies on the firm's permissible use of generative AI. Restating earlier opinions on outsourcing, it adds that lawyers relying on an outside provider must ensure it will "do the work capably and protect the confidentiality of information relating to the representation," and that those concepts also apply to AI providers and tools. For self-learning tools whose output could disclose client information, it says the client's informed consent is required before that information goes in.

In footnote 3 of its 17 February 2026 memorandum in United States v. Heppner, the Southern District of New York (Judge Rakoff) said that even if certain information the defendant entered into Claude was privileged, he waived the privilege by sharing it with Claude and Anthropic, as he would have by sharing it with any other third party. That is one district court, and it concerned a client using a public tool, not a firm's vendor; our analysis of the Heppner ruling sets out where the line falls.

The safer pattern mirrors the accounting one: develop on synthetic matters, deploy in the firm's own tenant, have counsel direct how the tool is used, and bar the provider by contract from training on or retaining firm data. The contract terms protect confidentiality; whether counsel's direction keeps the work inside the privilege is a possibility the memorandum raises at page 7 but did not decide.

Questions to settle before you decide

Is AI part of what we sell, or part of how we deliver it?
Who on our side will own each workflow after launch, by name?
Does the contract assign the code, prompts and evaluation set to us in writing?
Will the build team see client data, from which country, and under what consent or contract?
When the model is retired, who moves the agent, and inside what notice window?

If you cannot answer the second question, no sourcing model will save the project. If you can, hiring versus outsourcing becomes a question of capacity and timing, because control was settled when you named the owner.

Book a free AI assessment

Thirty minutes, no commitment. We map one workflow, make the build or buy call, and scope the smallest thing worth shipping.