Outsource AI Agent Development: A Buyer's Guide

Outsourcing AI agent development works well for most growing businesses, provided you outsource the building and keep the ownership. Keep four things on your side: the business decisions, the accounts and credentials, the test set that defines "correct", and the right to walk away with everything. Then choose a team on its production record and time-zone overlap, not on its hourly rate. This guide covers what to hand over, what it costs in the US, UK and offshore, what belongs in the contract, and how to try a team before you commit.

This post is part of our guide to rolling out agents, which starts at wizeb.com/blog/how-to-implement-ai-agents-in-business. If you have not decided whether to build a custom agent at all, start with wizeb.com/blog/build-vs-buy-ai-agents-2026-decision-framework.

The Problem: The Rate Is the Easy Part

The case for an outside team is reasonable. Fortune's report on MIT NANDA's GenAI Divide research (finance.yahoo.com/news/mit-report-95-generative-ai-105412686.html) says tools bought from specialist vendors and partnerships succeeded about 67% of the time, while internal builds succeeded about a third as often. Treat that as a signal, not a law: it draws on self-reported data and a small sample, and companies that buy may differ from those that build. We also sell this service, so read it with that in mind.

The trouble starts after the quote. An owner compares $40 an hour with $180 an hour, picks the lower number, and six months later has an agent that works in a demo, runs on the vendor's OpenAI account, and has prompts nobody else can find. The money lost is rarely the build fee. It is the months spent, plus the rebuild when the team moves on.

What to Outsource and What to Keep

Split the work by who can recover if it goes wrong. Engineering skill can be rented. Knowledge of your business and control of your systems cannot.

Hand to the outside team

• Agent architecture, model choice and prompt engineering.
• Integrations with your CRM, helpdesk, ERP or phone system.
• Retrieval over your documents, plus logging and monitoring.
• Running the test suite and fixing what fails.

Keep on your side

• One named owner with about three to five hours a week and the authority to make decisions.
• The definition of the job: what the agent does, what it never does, and when it hands off to a person.
• The test set. Your staff decide what the right answer is on real past cases, and the vendor does not mark its own homework.
• Every account: cloud, model provider, repository, phone numbers and API keys, in your company's name and on your card.

That last point costs nothing and keeps you from being held hostage. The vendor gets scoped access to your accounts. You never need to ask for access to theirs.

Onshore, Nearshore or Offshore?

Rates vary widely. A senior US or UK agency often bills $150 to $250 an hour (about £120 to £200), while experienced offshore teams commonly bill $30 to $70. Our pricing guide at wizeb.com/blog/ai-agent-cost-guide puts a single-task agent at $5,000 to $15,000 on blended rates, plus $100 to $600 a month to run. Eastern Europe and Latin America usually land between the two.

Agents are tuned through repeated test runs, so the hidden cost is the wait between a failed test and a fix. Compare a team with four hours of daily overlap, where you get same-day answers, to one with no overlap, where each question takes a day. Twenty rounds of questions become twenty lost days, whatever the rate. Offshore can still be the best value. Require the overlap in writing instead of assuming it.

• Choose onshore when the agent touches regulated data, or you need someone in the room with your staff.
• Choose nearshore or offshore when the job is clearly scoped, your owner is available, and the team can show live agents and give you at least three to four shared hours a day.
• Avoid any team, at any rate, that swaps its senior engineers for juniors after the sale. Name the people in the contract.

What to Put in the Contract

Much outsourcing regret could have been prevented by a clause. Have a lawyer review the final wording, but make sure these eight points are covered:

1. 1IP assignment. Paying for work does not always transfer ownership. Under US law, commissioned software often falls outside "work made for hire" (copyright.gov/circs/circ30.pdf), and UK contractors generally keep copyright unless they assign it. Get a written assignment covering code, prompts, workflows, test sets and documentation.
2. 2Acceptance criteria. Agree a pass mark on your own test set before the build starts, for example 95% correct on 300 past cases, with no serious errors on the cases you flag as high risk.
3. 3Payment tied to milestones. Pay for discovery, a passing test run, shadow mode and go-live, not for hours or calendar dates.
4. 4Data processing. Name the model providers, regions and retention periods. If EU or UK personal data leaves the country, you need a data processing agreement and a valid transfer mechanism, such as EU Standard Contractual Clauses or the UK IDTA.
5. 5Access rules. Separate service accounts with the narrowest permissions, no production data on personal laptops, and all access removed on the day the contract ends.
6. 6Named people and overlap hours. Name the lead engineer, the replacement process and the shared hours each day.
7. 7Running cost and model changes. Require an estimate per conversation or per task, plus a rule for who pays to retest when a model is replaced.
8. 8Exit. Define a handover period, a documented runbook, and a test that another engineer can deploy the agent from your repository without the vendor's help.

For the wider vendor questions, use the ten in wizeb.com/blog/how-to-choose-an-ai-agent-development-company. They cover production references and red flags, which this list does not repeat.

Try the Team Before You Commit

Do not start with a six-month engagement. Buy a small paid project, typically two to three weeks and $2,000 to $5,000 depending on rates. Give the team 100 real, anonymized cases from one job and ask for three things: a working prototype scored against those cases, a list of what failed and why, and a fixed quote for the production build.

You learn more from how they work than from the prototype. Did they ask about your exceptions or only your happy path? Did they report failures without being pushed? Did the engineer you met do the work? A team that tells you it scored 71% and explains the gap is worth more than one that sends a polished demo with no numbers.

How to Manage an Outsourced AI Agent Build

• Hold a 30-minute weekly review of the accuracy score, not a status update. The number going up is the only progress that counts.
• Keep all code in your repository from day one, with changes visible as they are made.
• Have one of your staff review a sample of outputs every week and log corrections where the team can see them.
• Roll out in stages: shadow mode, then draft-and-approve, then spot checks. Our timeline guide at wizeb.com/blog/ai-agent-implementation-timeline shows how long each stage usually takes.
• Budget for life after launch. A common rule of thumb is 15 to 25% of the build cost per year, whether that goes to the same team, a new team or your own.

Example: Two Ways to Outsource One Agent

Here is a hypothetical example. A 60-person UK distributor wants an agent to answer order-status emails, which take up about 25 staff hours a week. It first talks to a low-cost team that offers a fixed $6,000 build on its own hosted platform, with no test set and no mention of data transfer. The price looks good, but the distributor would own nothing and could not leave without a rebuild.

It chooses an offshore team at a higher rate that agrees to four shared hours a day. The team runs a $3,000 trial on 100 past emails and scores 78%, finding that most misses involve split shipments. The $14,000 production build runs in the distributor's own cloud account, with payment tied to a 95% pass on 300 cases. It goes live in draft-and-approve mode in week seven. Because the repository, keys and runbook are all in the distributor's name, it can move support to another team in year two if it wants to, and that option also keeps the original team sharp.

How WizeB Fits

WizeB builds custom AI agents, workflow automations and voice agents for growing businesses in the US, UK and Europe, and we built this checklist to be used on us. We scope one job first, test against your own past cases, run shadow and draft-and-approve stages before the agent works alone, and hand over the prompts, workflows, test sets and documentation. We also show you the monthly running cost before you sign. See how we work at wizeb.com/services/ai-agents.