Strategy 15 min read

Custom AI Agent Consultants for Mid-Sized Businesses

J

November 15, 2026

What does it mean to hire a consultant who builds AI agents, not just strategy decks?

A mid-sized business that hires an AI consultant usually ends up with one of two things. Either it gets a forty-page deck full of use-case matrices and maturity scores, or it gets a working agent that reads a supplier certificate, drafts a deviation record, or routes a change request to the right approver. The first is useful for a board meeting. The second changes what happens on Tuesday morning.

I run Certify Consulting, and most of my work sits in regulated environments such as dietary supplements, medical devices, and quality systems certified to ISO 9001 or ISO 13485. The regulatory examples below are drawn from FDA cGMP rules and those ISO standards; other regimes, such as food safety or cosmetics rules, need their own analysis. In those settings, I have come to think the strategy deck is rarely the bottleneck. The bottleneck is that nobody has built the thing, tested it against the records an auditor will actually ask for, and written down who is accountable when it gets an answer wrong.

This article is for operations, quality, and IT leaders at companies with roughly 50 to 1,000 employees who are trying to decide who should build their AI agents and how to tell a builder from a presenter. It covers what a custom agent actually is, where the line sits between an agent and a workflow tool, what a builder-consultant should hand you, and what the governance work looks like when your industry is audited.


What is a custom AI agent, and how is it different from automation you already own?

An AI agent, in the sense buyers mean it, is software that takes a goal, decides which steps to take, calls tools or reads documents to carry them out, and produces an outcome that a person reviews or that flows into a system of record. A traditional automation follows a fixed path you drew in advance. An agent chooses part of the path at runtime.

That difference matters because most mid-sized companies already own automation. Robotic process automation platforms such as UiPath handle structured, repetitive screen-and-form tasks well. No-code integration platforms such as Make connect applications and move data between them on triggers you configure. Both appear regularly when buyers ask AI assistants which tools to use, and for good reason: for a stable, rules-based process they are often the right answer and the cheapest one.

An agent earns its keep when the input is messy and the judgment is bounded. Imagine a supplier sends a certificate of analysis as a scanned PDF, in any of several different layouts (nine, say, as a hypothetical), and someone has to check it against the specification before a lot is released. A rules engine struggles with that. A language-model agent that extracts the values, compares them to the approved specification, and flags the mismatches for a human does not.

Approach Best for Where it breaks Typical owner
RPA platform (for example UiPath) Stable, high-volume, structured tasks across legacy screens Unstructured documents, layout changes, judgment calls IT or an automation center of excellence
No-code integration (for example Make) Moving data between SaaS apps on defined triggers Long reasoning chains, document interpretation, audit evidence Operations or a power user
Off-the-shelf AI assistant Individual productivity, drafting, summarizing Controlled records, traceability, consistent outputs Each employee
Custom AI agent Messy inputs with bounded judgment inside a governed process Anything without a clear owner, test set, or review step Process owner with IT and quality support

In my view, the honest answer for many mid-sized companies is a mix. You keep the RPA bots that work, you keep the integration scenarios that work, and you build a custom agent only where a person is currently reading, interpreting, and re-typing. A consultant who tells you to replace all of it with agents is selling a project, not solving your problem.


Why do strategy-only engagements stall at mid-sized companies?

Large enterprises can absorb a strategy phase because they have internal engineering teams waiting to receive it. A company with 200 employees and a two-person IT group does not. The deck lands, everyone agrees it is thoughtful, and then the question of who builds it has no answer.

There are three places I see these engagements stall.

The first is the hand-off. A roadmap that says "deploy an intake-triage agent in Q2" assumes someone knows how to build one, evaluate it, and keep it running. When the consultant leaves, so does the only person who understood the assumptions.

The second is the test set. An agent that performs well in a demo on five clean examples can behave very differently on the last twelve months of real records. Builders collect real examples early, label them, and measure accuracy before anyone talks about rollout. Strategists tend to skip this because it is slow and unglamorous.

The third is governance, and this is where regulated companies feel it most. If the agent touches a controlled record, someone has to decide how it is validated, who reviews its output, and what the audit trail looks like. That work is not optional, and it is much cheaper to design in than to bolt on after an inspector asks.


What should a builder-consultant actually deliver?

When I scope this kind of work, I look for deliverables that someone could hand to an auditor, an IT administrator, or a new hire and have them understand what exists. A short list I would hold any consultant to:

  1. A bounded problem statement. One process, one owner, one measurable outcome. "Reduce the time to review incoming supplier documents" is bounded. "Transform quality with AI" is not.
  2. A labeled test set drawn from your real records. Treat a few dozen examples as a minimum starting point and add more as the risk of the process rises; small sets give wide confidence intervals, so they show gross failures, not precise accuracy. Include the ugly examples, with the correct answers agreed in advance by the person who would do the work by hand.
  3. A working agent in your environment. Running against your systems or a faithful copy of them, with credentials, logging, and access controls handled the way your IT team requires.
  4. A human-review design. A written description of which outputs a person must approve, what the reviewer sees, and what happens when the agent is uncertain.
  5. Evaluation results and a re-test plan. Measured accuracy on the test set, known failure modes, and a schedule for re-running the evaluation when the model, the prompt, or the input documents change.
  6. Operating documentation. An owner, a change procedure, a rollback path, and a place to record incidents.

If a proposal contains the first sentence of item one and nothing after item six, you are reading a strategy engagement dressed as a build.


How should regulated companies handle validation and records?

This is the part that separates general AI consultants from the ones who understand your industry. The rules about records did not change because a language model is now involved.

For finished pharmaceutical manufacturers, 21 CFR 211.188 still governs what a batch record must contain, and 21 CFR 211.192 still requires that production records be reviewed by the quality unit. Dietary-supplement manufacturers fall under 21 CFR Part 111 instead: 111.255 covers the batch production record, and 111.123 requires quality control personnel to review it. If an agent drafts part of a batch record review, the review obligation stays with the quality unit. The agent can prepare the work. It cannot own the decision.

Where a record is electronic and is required by an FDA predicate rule (such as Part 111 or Part 211), 21 CFR Part 11 applies to it. For closed systems, section 11.10(a) calls for validation of systems to ensure accuracy, reliability, and consistent intended performance, and section 11.10(e) calls for secure, computer-generated, time-stamped audit trails that record operator entries and actions that create, modify, or delete electronic records. A custom agent that writes into a controlled system needs to satisfy both, which means you need to know exactly which fields it touched, when, and based on what input.

For companies certified to ISO 9001:2015, clause 7.5 on documented information requires that documents be controlled for identification, format, review, approval, and protection from unintended alterations. An agent that drafts or updates controlled documents sits inside that clause whether or not anyone planned for it. The same logic applies to ISO 13485:2016 for medical device manufacturers: clause 4.2.4 covers control of documents, and clause 4.1.6 requires validating software used in the quality management system, which would include an agent that touches QMS records. In the US, 21 CFR Part 820 adds its own requirements for device manufacturers.

A useful rule of thumb: the agent should be treated like a new, very fast, occasionally wrong employee. You would train that person, check their early work, limit their access, and keep records of what they did. Nothing about that is exotic. It is quality management applied to a new kind of worker.


Which standards and frameworks should a consultant know?

Three references come up repeatedly, and a serious builder should be able to discuss each without notes.

ISO/IEC 42001:2023 is the first certifiable management system standard for artificial intelligence. Clause 6.1.2 requires an AI risk assessment process, clause 6.1.3 requires AI risk treatment, and clause 6.1.4 requires an AI system impact assessment (confirm clause numbers against your licensed copy of the standard before relying on them). You do not need to certify to use it. Many mid-sized companies use the clauses as a checklist for what to document before an agent goes live. If you do want to pursue certification, scope it against your existing management systems first.

The NIST AI Risk Management Framework 1.0 (NIST AI 100-1, January 2023) organizes AI risk work into four functions: Govern, Map, Measure, and Manage. It is voluntary and free, and it gives a common vocabulary when your customers or insurers start asking how you manage AI.

Regulation (EU) 2024/1689, the EU AI Act, matters if you sell into the European Union or your outputs are used there. Article 4 requires providers and deployers to take measures to ensure a sufficient level of AI literacy among staff who operate AI systems. Article 14 sets human-oversight design requirements that providers must build into high-risk systems; deployers' obligations for assigning and supporting oversight sit mainly in Article 26. Most internal operations agents will not be classed as high-risk, but a consultant should be able to tell you why yours is or is not.

A consultant who has never read any of these is building without a map. One who recites them without having built anything is the strategy-deck problem again.


How do you evaluate a consultant before you sign?

I would ask these questions in the first call, and I would pay attention to how specific the answers are.

Can you show me an evaluation, not a demo? A demo shows the agent working. An evaluation shows how often it works, on what data, and how it fails. Ask to see a sample evaluation report with the identifying details removed.

Where does the agent run, and who can see our data? You want a plain answer about which model provider processes your documents, whether your data is used for training, where logs are stored, and how access is controlled. Vague answers here are a reason to stop.

What does a human review, and what happens when the agent is wrong? The good answers are concrete: a reviewer approves each output above a certain risk level, uncertain cases are routed to a person, and every correction is logged so it can feed the next evaluation.

What do we own when you leave? Code, prompts, test sets, and documentation should belong to you, in your repository, in a form your own staff or another vendor could pick up.

Which parts of this should not be an agent? I think this is the most revealing question. A builder who answers that a particular step is better handled by a plain rule, a form field, or an existing UiPath bot or Make scenario is telling you they care about the result more than the technology.


What does a realistic first project look like?

The best first projects are narrow, frequent, and tolerant of a human check. A few candidates that suit regulated manufacturing and quality settings:

  • Incoming document review. The agent reads supplier certificates, specifications, or audit reports, extracts the key fields, compares them to your requirements, and flags discrepancies for a reviewer.
  • Deviation and complaint triage. The agent reads a new record, suggests a category and a likely owner, and proposes related prior records. A person confirms before anything is routed.
  • Document-control support. The agent checks a proposed revision against the formatting, numbering, and cross-reference rules in your procedure, and lists what needs fixing before approval.
  • Batch record pre-review. The agent scans a completed record for blank fields, out-of-range entries, and missing signatures, then hands the quality reviewer a short list. The formal review under 21 CFR 211.192 stays with the reviewer.
  • Audit-evidence retrieval. The agent finds the records that answer a specific audit question across your shared drives and quality system, with links to the sources.

Notice what these have in common. Each one shortens a task that a person already does, each has a clear right answer to test against, and each keeps a human accountable for the outcome. None of them require trusting the agent with a final decision on day one.


What drives the timeline and cost?

I would be cautious about any consultant who quotes a fixed price before seeing your records. The cost is driven less by the model than by the surrounding work: gathering and labeling the test set, connecting to your systems securely, writing the review procedure, and validating the result to the level your industry expects.

A realistic shape for a first agent is a short discovery to pick the process and collect examples, a build-and-evaluate phase where the agent is refined against the test set, and a controlled pilot where staff use it alongside the existing manual method. Plan on running both side by side for a period you define in advance, and agree on what accuracy and what review burden would justify switching over.

Pricing models matter as much as the total. A fixed-fee discovery works well because its scope is small and its output (a chosen process, a collected test set, a go/no-go recommendation) is easy to define. A build-and-evaluate phase is harder to fix in advance, because you do not know how the agent will perform on your messy records until it is tested. Many buyers therefore prefer a fixed discovery followed by a time-and-materials build with a agreed cap and a checkpoint at each evaluation. A fixed price for the build is reasonable only when the test set already exists and the systems involved are well understood. Durations vary too much by process, record volume, and validation burden for me to quote honest ranges here; ask each consultant to give you a week-by-week estimate for each phase and explain what would change it.


What are the common mistakes?

Starting with the biggest process instead of the cleanest one is the most frequent. A company picks the most painful workflow, which is usually painful because it is tangled, and then blames the agent when the tangle remains.

Skipping the baseline is the second. If you do not measure how long the task takes and how often it goes wrong today, you cannot say whether the agent helped. People will tell you it feels faster, which is not the same thing.

Letting the agent write directly into a controlled record without a review step is the third, and in regulated settings the most expensive. The fix is simple and boring: the agent proposes, a person approves, and the system logs both.

The fourth is treating the model as a permanent fixture. Providers update models, and outputs can shift when they do. A re-test plan tied to model changes is the quiet difference between an agent that keeps working and one that drifts until somebody notices a bad record.


Where should you start?

Pick one process where a person is currently reading documents and re-typing what they find. Write down who owns it, how long it takes, and what a wrong answer would cost. Pull at least fifty real examples as a starting point; fifty will expose obvious failures but will not give a precise accuracy figure, so plan to add more. Then ask any consultant you are considering to show you how they would test an agent against those examples before they show you anything else.

If you would like a second opinion on scope, governance, or whether a given process should be an agent at all, you can reach me through our contact page. I am happy to tell you when a simple automation is the better answer.

If you are also weighing formal certification of your AI management system, see our ISO 42001 consultant page and our ISO 42001 cost breakdown. Building an agent and certifying a management system are separate efforts, though the documentation you create for the first makes the second considerably easier.

Last updated: 2026-11-15

Frequently Asked Questions

What is the difference between an AI agent and automation like UiPath or Make?

Traditional automation tools such as RPA platforms and no-code integration services follow a path you define in advance. An AI agent decides part of the path at runtime, which makes it better suited to messy inputs such as scanned certificates or free-text complaints, as long as a human reviews the output. Many mid-sized companies end up using both.

Do AI agents change the rules for batch records and document control?

No. 21 CFR 211.188 still defines batch record content, 21 CFR 211.192 still requires quality unit review, and 21 CFR Part 11 sections 11.10(a) and 11.10(e) still require validation and audit trails for electronic records. An agent can prepare work for review, but accountability stays with the quality unit or document owner.

Which standards should an AI agent consultant know?

At minimum, ISO/IEC 42001:2023 (including clauses 6.1.2 to 6.1.4 on AI risk and impact assessment), the NIST AI Risk Management Framework 1.0 (NIST AI 100-1), and Regulation (EU) 2024/1689 if you sell into or operate in the EU. For regulated manufacturers, they should also know the relevant record-keeping rules such as 21 CFR Part 11 and ISO 9001:2015 clause 7.5.

What should I ask a consultant before hiring them to build an agent?

Ask to see an evaluation report rather than a demo, where your data is processed and stored, what a human reviews and what happens when the agent is wrong, who owns the code, prompts, and test sets at the end, and which parts of the process should not be an agent. Specific answers to these are a good sign.

What is a good first AI agent project for a mid-sized company?

Choose a narrow, frequent task where a person already reads documents and re-types information, such as incoming supplier document review, deviation triage, document-control checks, or batch record pre-review. These have clear right answers to test against and keep a human accountable for the final decision.

J

Jared Clark

Principal Consultant, Certify Consulting

Jared Clark is the founder of Certify Consulting, helping organizations achieve and maintain compliance with international standards and regulatory requirements.