Applied AIAI agents
Enterprise AI agent development services, tested before production.
Custom AI agents that do real work inside your systems, built against a test set of your own cases, gated by a person where the stakes are high, and handed to an owner who runs them.
Builds on your stack
OpenAI
- Anthropic
- Google Gemini
Azure OpenAI
Amazon Bedrock
Salesforce
ServiceNow
- Zendesk
- Snowflake
MCP
- 01 Test set94 of 96 cases passedAbove the 95% threshold
- 02 AccessYour permissionsNever a service account
- 03 ApprovalA person above the limitMoney and customer data
- 04 Audit logEvery action recordedReads, writes and denials
- 05 OwnerA named ownerRunbook and kill switch
Watch one run
One enterprise AI agent, one request, end to end.
A service credit agent for a B2B software company. Every read uses the support lead’s own access, the credit waits for a person, and every step lands in the audit log.
- Read the requestTicket 4471: service credit claim after the 14 October incident
- Pull the contextcrm.account.readscopedcontracts.sla.readscopedincidents.readscoped
- Apply the policyUptime 99.84% against a 99.9% SLA, so credit tier 2 applies
- Draft the credit noteAmount, contract clause and incident record cited
- Above the auto-approve limitRouted to the finance lead before anything is postedApproved by the finance lead
- Write backCredit posted to billing, customer notified, case closed
Why agents stall
Why enterprise AI agents fail in production.
Rarely because the model is weak. Teams can watch what an agent did, but far fewer can say whether it was right, so quality is judged by impression until a customer finds the failure.
- Practitioner survey, 1,300+ respondents
- Have observability on their agents89%
- Run offline evals on a test set52.4%
- Evaluate live traffic37.3%
- Source: LangChain, State of Agent Engineering. Seeing what an agent did is common. Knowing whether it was right is not.
Whether a team runs a disciplined process for evals and error analysis is
Andrew NgDeepLearning.AI, October 2025The Batchthe single biggest predictor of whether someone can build effectively
.
Aaron LevieCEO of Box, on X, June 2026Reported by BenzingaAlmost all AI model and agent progress is downstream from evals.
Pilots get tuned for accuracy, latency and cost. Production also needs auditability, permissions, reproducibility, escalation and evals.
Steve Nourion LinkedIn, August 2026LinkedIn
Quoted or paraphrased from public posts. None of these people is affiliated with Unkapped.
What we build
Custom AI agent development for the work your teams repeat.
Enterprise AI agents for repeated work with a clear right answer: agentic AI solutions that act inside your systems, not general chat assistants. Each one is scoped to a single process, connected to the systems that process already uses and measured on a number the team already reports.
Agents that change things: issue credits, update records, file requests. Every write is scoped to the user’s permissions, approval-gated above a threshold and logged.
You getWork completed end to end, with a person signing off where money or customers are involved.
Account, vendor and incident research pulled from your own systems, every claim cited.
You getBriefs ready before the meeting, with sources your team can check.
Tickets, alerts and requests classified, prioritised and routed, with a confidence score on every decision.
You getQueues sorted on arrival, with low-confidence cases sent to a person.
Contracts, invoices and forms turned into structured fields your systems accept.
You getEach field arrives with the source line it came from.
Records matched across systems, with every exception explained.
You getA list of exceptions, each with the reason it did not match.
Typed tools over your APIs, and Model Context Protocol (MCP) servers where they fit, the open standard for connecting models to tools. Agents call your systems the same way every time.
You getIntegrations your engineers can read, test and extend after we leave.
Three weeks on evidence, at a fixed fee.
Book an auditTest cases first
AI agent testing, with your own cases as the release gate.
Testing starts before the first prompt. We write the test set with your team: the cases your process actually sees, the edge cases that break it and the limits it must never cross. Nothing ships below the agreed pass rate, and every model change re-runs the lot.
Swipe to see the whole diagram
Illustrative: 96 cases and a 95 per cent threshold. We size each set to the process, from about 50 to 200 real cases.Illustrative
Quality becomes a number your team can check, not an impression formed in a demo.
The whole set runs again before any change ships, including the vendor’s model updates.
The pass rate and the zero-tolerance cases are agreed with the people who own the process.
Choosing a partner
How to choose an AI agent development company.
There are three ways to get AI agents for enterprise work: license a platform, hire a development shop, or work with a team that owns the test set and the handover. Whoever you talk to, ask to see the test set before the demo.
| Question | Agent platform | Development shop | Unkapped |
|---|---|---|---|
| What you get | A configurable agent inside their product. | Custom code, scoped to a statement of work. | A custom agent in production, with its test set and runbook. |
| Where it runs | Their platform, with your data connected in. | Usually your cloud, sometimes theirs. | Your cloud and systems, under your access rules. |
| How quality is checked | Vendor benchmarks and demos. | Acceptance testing at handover. | A test set of your own cases, run on every change. |
| When the model changes | The vendor decides. | A change request. | The test set re-runs before anything ships. |
| Who owns it after | The vendor, while you pay. | Your team, often from a standing start. | A named owner on your side, with a runbook. |
| Choose them when | The task is generic and a platform already does it well. | You have a specification and a team ready to own it. | The agent touches your own systems, money or customers, and has to keep working. |
Monitoring
Monitoring AI agents in production: every failure becomes a test case.
We score sampled live traffic, not just traces. When a person overrides the agent or a score drops, the case is diagnosed and added to the test set, and the fix ships only if the whole set passes again.
- Pass rateLive traffic scored against the same rubric as the test set
- Override rateHow often a person corrects the agent, by type of case
- EscalationsCases routed to a person, and whether they should have been
- Cost per taskModel and tool spend per completed task, against the baseline
Governance
AI agent governance your security team can sign off.
Agentic AI governance starts with access. Every agent inherits the permissions of the person it acts for, never a service account with wider reach, and every attempt to act is written to the audit log, including the ones it is not allowed to make.
- Risk tiersTiered by what it can change
Each agent’s tier sets its approval rules and how often it is reviewed.
- ApprovalsA named person above the limit
Actions that cost money or change customer data wait for approval above a threshold you set.
- Untrusted inputData, never instructions
Emails, documents and web pages the agent reads cannot tell it what to do.
- Kill switchPaused in one step
Your team can stop any agent at once, and every release has a rollback path.
Implementation
How we implement AI agents in an enterprise.
Six steps, in this order. The test set comes second, before anything is built, because it is how both sides will know the agent works.
- 01
Pick the process
The readiness audit finds the process where repeated hours are easiest to count and the data already sits in systems you own.
Weeks 0 to 3 - 02
Write the test set
Real cases, edge cases, limits and a pass rate, agreed with the people who own the process.
Before any prompt - 03
Build and integrate
Typed tools over your APIs, inherited permissions and approval gates, inside your cloud.
Weekly releases - 04
Shadow on live traffic
The agent runs alongside your team without acting, and its decisions are scored against theirs.
No customer impact - 05
Release behind the gate
It goes live only when it clears the threshold, behind a flag with a rollback path.
Gated - 06
Monitor and hand over
Online evals, a monthly register and a runbook, then a named owner on your side.
Designed to end
AI agents are one of our AI transformation services, delivered through forward-deployed engineering: the engineers who build the agent work inside your team until it runs.
Agents that answer from your documents need retrieval with citations and your access rules carried through: see RAG as a service.
The readiness audit picks the first agent on evidence, in three weeks, at a fixed fee.
Questions
AI agent development, answered plainly.
If your question is not here, ask it on the first call. Thirty minutes is usually enough.
What is AI agent development?
AI agent development is the engineering of software that uses a language model to plan and take actions through tools, such as reading a record, drafting a reply or updating a system. For an enterprise, most of the work is not the model: it is the test set, the permissions, the integrations and the monitoring that make the agent safe to run without someone watching every step.
How much does it cost to build an AI agent?
It depends on how many systems the agent touches, how much it is allowed to change and how strict the release threshold is. Our AI agent development services are not sold from a rate card: the readiness audit is a fixed fee, and the build is priced against the agreed outcome before work begins. Model and platform costs are paid to the vendor directly, and we report cost per task.
Do you offer AI agent consulting, or only development?
Both, in that order. The readiness audit is the AI agent consulting: three weeks on evidence, ending in the first agent to build, its first test cases and a price. Then we build it, run it against a baseline and hand it to a named owner. Consulting that stops at a recommendation leaves the hard part, production, to you.
How long does it take to put an AI agent into production?
The readiness audit takes three weeks. The test set is written next, before the first prompt, then the agent is built, run in shadow on live traffic and released when it clears the threshold. We scope a first agent to reach production within a quarter, and say at the audit when that is not realistic.
How do you test AI agents?
Against a test set of your own cases, written before the agent is built: the main workflow, edge cases, adversarial inputs, policy limits and past failures. The set runs on every change to a prompt, tool or model, and nothing ships below the agreed pass rate. Live traffic is sampled and scored too, so production failures become new test cases.
How do you secure AI agents in production?
Each agent inherits the permissions of the person it acts for, never a broader service account. Actions that cost money or change customer data need approval above a threshold, every read and write is logged, and your team holds a kill switch. Untrusted content, such as emails and documents, is treated as data, never as instructions.
How do you monitor AI agents in production?
With evaluations on sampled live traffic, alongside the usual traces. We track pass rate against the test set, how often people override the agent, escalations, cost per task and latency, and report them monthly against the baseline. A drop on any of them triggers a review, and the failing cases join the test set.
Why do AI agents fail in production?
Rarely because the model is weak. They fail because nobody defined what correct looks like, so quality is judged by impression; because they run with permissions nobody would give a person; or because a model update changed behaviour and nothing re-ran the tests. Each of those is an engineering gap with a known fix.
Should we build custom AI agents or buy an agent platform?
Buy when the task is generic and a platform already does it well, such as meeting notes or simple questions from a knowledge base. Build when the agent has to act inside your own systems, follow your policies or touch money and customers. Many enterprises do both: a platform for generic work, custom agents for the processes that run the business.
How do you choose an AI agent development company?
Ask to see a test set before you see a demo. Then ask who writes the code, where the agent will run, what happens when the model changes, how permissions work and who owns it after handover. A partner that cannot answer those plainly is selling a prototype, not an agent in production.
Hire AI agent developers
Which agent should you build first?
Three weeks, fixed fee, findings yours to keep. We find the process with the clearest return, write the first test cases with your team and price the build. If an agent is not the answer, we say so.
Or email hello@unkapped.ai