AI enablement · Agents and process automation
When an AI agent is the right choice, and when it is not
We automate multi-step processes, with carefully designed tools, saved progress, failure handling and points where a person approves the next step. We also tell you when a simple fixed workflow would be the better choice, which is more often than the market suggests.
Where this starts
Giving AI more freedom has a cost
An AI agent that decides its own steps is harder to test, debug and explain than a workflow with fixed steps. Sometimes that cost is worth paying, because the steps genuinely differ from case to case. Often it is not. The team has then taken on the hardest possible way of solving a problem that a simple fixed workflow could have solved in a couple of weeks.
What we usually find
- An AI agent used where the steps were always going to be the same
- An agent stuck in a loop, repeating steps, spending money and achieving nothing
- Failures that cannot be reproduced, and so cannot be fixed
- No record of why the agent made the choices it made
- A process that stopped halfway, with no way to resume it or undo what was done
- A cost per run that nobody noticed until the invoice arrived
Our position
We often recommend a fixed workflow with a single AI step inside it. It sounds less impressive, but it is usually the right answer.
What the work covers
Designing processes that automation can complete
The key engineering decisions are about limits: what the agent may do on its own, what it must ask permission for, and what happens when a step fails after earlier steps have already taken effect.
Choosing the right approach
The options are a fixed workflow, a workflow with some AI steps, or a genuine AI agent. We choose based on how much the steps really vary in your cases, not on current industry trends.
Designing the tools
Each action the agent can take is a narrow, clearly described tool that checks its inputs. Most agent failures are really caused by a tool that was too broad or too vaguely described to be used correctly.
Saving progress
Progress is saved, so a long process survives a restart and can be resumed or undone. It does not have to start again from the beginning, with whatever the first attempt left behind.
Human approval
A person must approve anything that cannot be undone, and is shown the information needed to decide quickly. If approvals are slow, they stop being a safeguard and become a bottleneck.
Handling failures
We plan in advance what happens when step four fails after step three has already sent an email. Ways to reverse or correct earlier steps are designed from the start, not worked out during the first real incident.
Limits on steps and spending
Each run has a firm limit on steps, usage and cost. An agent stuck in a loop is a financial problem as well as a technical one, and it tends to happen at the weekend.
How each run is accountable
Every run can be explained
If nobody can check what an AI agent did afterwards, it will not be trusted with anything important, and it should not be.
- Trace
- Every step, tool used, input and decision is recorded together. A run can be replayed and understood, rather than pieced together from guesswork.
- Limits
- Limits on steps, usage and spending are enforced for every run. When a limit is reached, the run is stopped and reported, not quietly allowed to continue.
- No duplicates
- Actions are safe to repeat. Recovering from a failure does not send a second copy of an email, order or payment that the first attempt already sent.
- Approvals
- Each approval records who approved it and what they were shown. This is what makes the automation defensible months later.
- Regression tests
- A fixed set of scenarios is re-run after every change. A small change to a prompt can change the agent’s behaviour several steps later in ways that ordinary tests will not catch.
What you are left with
- An honest recommendation on whether you need an AI agent at all
- Narrow, documented tools that check their inputs
- Processes that resume after a failure rather than starting again
- Limits on steps and spending for every run
- A full record of every run that can be replayed
Describe your process from start to finish
If the steps turn out to be the same every time, we will tell you, and build you something simpler.