Insights
What does an agent readiness review cover?
By Rare Agentics engineering teamLast updated First published 9 minute read
An agent readiness review looks at how work actually moves through your business and finds the tasks where an AI agent can safely take on work. Each candidate task is scored for value, frequency, risk and data readiness, and the shortlist comes with the bounds each agent would work within, what would always wait for a person, what data would leave your systems, and a costed roadmap.
Why start with a review instead of a pilot?
Most businesses that want to "do something with AI" start with a pilot. Someone picks a task that sounds impressive, a supplier builds a demo, and three months later nobody can say whether it saved anything or what it would take to trust it with real work.
The problem is rarely the model. It is the choice of task. A pilot chosen for how it looks in a board meeting tends to be the wrong first agent: too broad, too hard to check, or sitting on data that is scattered across inboxes and spreadsheets.
A review reverses the order. You look at the work first, decide which tasks are worth handing to an agent and which are not, agree the controls each would need, and only then spend money building anything. It adds two weeks at the start, and removes the months a badly chosen pilot can waste.
It also gives you something a pilot does not: a written answer you can show a board, an investor or a regulator about how AI is being adopted and what stops it going wrong.
What does an agent readiness review actually look at?
A good review covers five things. If a supplier's version leaves any of them out, ask why.
How the work is really done. Not the process diagram, the actual steps. That means interviews with the people who do the work and a walk through the systems they use. The interesting findings are nearly always in the gap between the official process and the workaround someone built in a spreadsheet three years ago.
The data each task touches. Where it lives, who owns it, how clean it is, and whether an agent could read it through a proper connection or only by someone exporting a file. A task with good value and bad data is a data project first and an agent project second.
The risk each task carries. What happens if the agent gets it wrong, how quickly anyone would notice, and whether the mistake can be undone. Drafting an internal summary and paying a supplier are not the same kind of task, even if the steps look similar.
The controls each task would need. What the agent may read, what it may do, and what always waits for a person. These should be specific enough to build against, not a paragraph of principles.
What data would leave your systems. Which AI providers would see which data, and what each provider keeps. This is the question legal and compliance teams will ask first, so it belongs in the review rather than in a panic after launch.
The output is a ranked shortlist and a costed roadmap. Our own Agent Readiness Review runs over two weeks: week one is interviews, systems walkthroughs and a review of the data each task touches; week two is scoring, controls for each shortlisted task, the roadmap and a read-out with your leadership.
How do you score which tasks an agent should take on?
Score every candidate task on the same four measures, so you are comparing like with like rather than arguing about whichever idea the most senior person likes.
| Measure | What you are asking | A strong candidate | A weak candidate |
|---|---|---|---|
| Value | What does doing this by hand cost in time, delay or errors? | Hours a week of skilled people doing repetitive steps | A few minutes a month |
| Frequency | How often does the task happen? | Many times a day or week | Quarterly, or only when something unusual happens |
| Risk | What happens if it is done wrong, and can it be undone? | A wrong result is easy to spot and cheap to correct | A wrong result moves money, deletes data or reaches a customer unchecked |
| Data readiness | Can an agent reach clean, complete inputs through a proper connection? | Structured records in one system with an API | Scattered across inboxes, PDFs and personal spreadsheets |
Two practical points on using the grid.
First, risk is not a reason to rule a task out. It is a reason to design the controls carefully. A high-value, high-risk task can still be a good candidate if the risky step waits for a person and everything before it is automated.
Second, data readiness is the measure most often scored too generously. People say "it is all in the CRM" and then the walkthrough shows half the records are in someone's notes. Score what you see in the systems, not what people believe is there.
What makes a task a good first agent, and what rules one out?
The best first agent is a task that is frequent, rule-bound and easy to check, where a person already does the same steps by hand. You want something narrow enough that everyone can tell whether it is working within a few weeks.
For example, an agent that reads supplier invoices, matches each to its purchase order and drafts the payment run for a person to approve. It happens often, the rules are clear, a mismatch is easy to spot, and the one consequential step (paying) waits for a person.
Other illustrative examples of the same shape:
- Triaging inbound requests into the right queue, with a reason attached that a person can check.
- Drafting replies to routine questions from your own documents, sent only after a person reads them.
- Classifying documents as they arrive and filing them against the right client record.
- Chasing overdue items from a list, with the message drafted and the send approved.
Signs that a task is a poor first choice:
- Nobody owns it. If no one person is accountable for the task today, no one will be accountable for the agent doing it.
- The rules live in someone's head. If the person doing it cannot explain how they decide, the first job is writing that down, not building an agent.
- Success is a matter of taste. Agents are easiest to trust where there is a right answer.
- It touches everything. "An agent that runs our operations" is a roadmap, not a task.
What controls should each shortlisted task have before an agent touches it?
For every task on the shortlist, the review should write down three lists: what the agent may read, what it may do, and what always waits for a person. These are agreed before anything is built, and they should be enforced by access controls, not by instructions in a prompt. An agent that has been told not to issue refunds but holds a key that can issue refunds is one bad input away from doing it.
Some things should always wait for a person: payments, deleting records, changes to customer records and releases to production. In our reviews they wait for a person at every level of autonomy, however good the track record.
It also helps to decide how much autonomy each task starts with. We use five levels of autonomy:
| Level | What the agent does | Typical use |
|---|---|---|
| Observe | Watches and records what it would have done, takes no action | Measuring accuracy against real cases before anything goes live |
| Suggest | Proposes an action for a person to take | Early weeks of a new agent |
| Act with approval | Prepares the action and carries it out once a person approves | Anything consequential, such as a payment run |
| Act and report | Acts on its own and tells a person what it did | Routine, reversible steps with a proven track record |
| Act without sign-off | Acts on its own within its bounds | Low-risk, high-volume steps where checking each one adds nothing |
Most tasks should start at Observe or Suggest. Bounds widen one level at a time, only when the business agrees and only on the evidence of how the agent has performed. The review should say where each task starts and what evidence would justify moving it up.
Finally, every agent action should be recorded: which agent acted, the sources it used, what it did and who approved it. An answer its sources do not support should go to a person rather than being sent. If a supplier cannot describe how that record works, the agent is not ready for real work. We cover the engineering side of this in more depth in how to keep AI agents away from production, and our FAQ on agents and controls sets out what we always enforce.
What data leaves your systems, and what does the provider keep?
An agent that uses a hosted model sends data to that provider. The review should say, for each shortlisted task, exactly what would be sent, to which provider, and what that provider stores and for how long under the terms you would be using.
The questions to ask:
- Which model provider would each task use, and why that one?
- Is any of the data personal, financial or commercially sensitive?
- Does the provider keep prompts or outputs, and for how long?
- Can the provider use your data to train its models under the terms you would sign?
- Could sensitive fields be removed or masked before they leave your systems?
The answers can change the shortlist. A task that looked ideal may need a different provider, an open model run in your own accounts, or a step that strips identifiers first. It is far cheaper to learn that in a review than after a build.
Where personal data is involved, the review should also say whether a data protection impact assessment is needed. UK GDPR requires one for processing likely to be high risk, and the ICO expects one for most uses of AI on personal data. It should also say where each provider processes the data, since many process it outside the UK.
We use Anthropic, OpenAI or open models, the best one for each job, and we check what each provider keeps before it touches your data.
What should you get at the end, and what should you ask a supplier?
A useful review leaves you with documents you can act on without the supplier in the room:
- A map of candidate tasks, each scored for value, frequency, risk and data readiness.
- A ranked shortlist, with the bounds each agent would work within and what would always wait for a person.
- What data would leave your systems, to which providers, and what they keep.
- A costed roadmap, with a clear first step.
If the roadmap only works if you hire the company that wrote it, it is a sales document. The review should stand on its own, so you can act on it with the author, with another partner or with your own team.
Questions worth asking any supplier before you commission one:
- Is the fee fixed and agreed in writing before the review starts?
- Will you interview the people who do the work, or only the leadership team?
- Will the controls be enforced by access controls, or by instructions to the model?
- Will the roadmap include costs, and will it be usable without you?
- What do you need from us, and how much of our people's time?
The time cost should be modest. For ours, we ask for a few hours with the people who do the work, read access to the systems we discuss, and one decision-maker for the read-out.
If the first task on the roadmap is clear, the natural next step is a small build to prove it. A proof sprint delivers one bounded feature, workflow or working prototype, live in your own accounts in 1 to 2 weeks, from £3,000 at a fixed price. A full custom AI agent for one defined task typically takes 3 to 6 weeks. For a sense of what that costs, see our guide to custom AI agent development costs in the UK.
How Rare Agentics can help
Our Agent Readiness Review is a two-week review of where agents can safely take on work across your business, at a fixed fee agreed in writing before it starts. You get the scored shortlist, the controls for each task, where your data would go and a costed roadmap, and you are free to act on it with us, with another partner or on your own. Ask us for the fee, which is fixed in writing before the review starts: get in touch. After launch, a Run and improve plan keeps each agent tuned as your business changes.