The first automation an organisation chooses does more than save time in one workflow. It sets expectations for everything that follows. Choose well and the organisation learns how to own, measure and govern automated work. Choose badly — too ambitious, too risky, or too trivial to matter — and the next proposal meets scepticism. This playbook gives a short decision framework built on four lenses — value, feasibility, risk and organisational readiness — applied in five steps: build a long list, screen out the non-starters, score the shortlist, shape the mission, and decide.
The problem: the first choice sets the tone
First automations usually fail in one of three ways. They aim too high: an end-to-end process with many systems and judgement calls, which takes months and never quite works. They aim at the wrong risk: an agent that talks to customers about money before anyone knows how it behaves. Or they aim too low: a task so small that its success proves nothing, and nobody notices when it is done.
The right first automation is the smallest piece of work that can deliver a meaningful, measurable result, with risk you can contain while you learn.
The four lenses
| Lens | The question | Good signs | Warning signs |
|---|---|---|---|
| Value | Does it matter if this improves? | Volume, waiting time customers feel, skilled people doing routine work | Nobody can say what it costs today |
| Feasibility | Can it be done with the systems and data you have? | Clear rules, accessible data, few systems | Data in people's heads; many undocumented exceptions |
| Risk | What happens when it goes wrong? | Reversible actions; a person can review | Irreversible, financial or external actions with no review |
| Readiness | Will the organisation adopt and own it? | A named owner who wants it fixed; a team willing to change | Owned by "the team"; the people doing the work were not asked |
Build a long list
Collect candidates from where the pain is, not from a vendor's list of use cases.
- Queues where work waits: shared inboxes, request forms, trackers.
- Work that skilled people describe as copying, chasing or checking.
- Complaints from customers about response time or consistency.
- Places where the same data is entered into more than one system.
- The question worth asking every team: what work consumes skilled time that you would rather a governed agent owned?
Aim for ten to twenty candidates, each described in one sentence with its owner. The cost of an unowned workflow describes a one-week queue audit that produces most of this list.
Screen out the non-starters
Remove candidates that fail a basic test before scoring the rest:
- No owner. Nobody is willing to be accountable for the outcome.
- No measure. You cannot say how it performs today.
- Undesigned process. People cannot describe the steps and decisions consistently. Design it first.
- High-stakes autonomy from day one. The value depends on an agent acting alone on irreversible, financial or customer-facing decisions.
What remains should be workflows with meaningful volume, clear rules, accessible data and a place for human review.
Score the shortlist
Score each remaining candidate from 1 to 5 on each lens. For risk, a higher score means lower risk.
| Score | Value | Feasibility | Risk (inverted) | Readiness |
|---|---|---|---|---|
| 5 | Large, visible, measurable gain | Rules clear; data in one or two systems | Fully reversible; low impact | Owner eager; team asked for it |
| 3 | Useful gain, moderately visible | Some rules unclear; three or four systems | Reversible with effort; review possible | Owner willing; team neutral |
| 1 | Marginal or unmeasurable | Rules unwritten; data scattered or inaccessible | Irreversible or external with no review | No owner; team opposed |
Two rules matter more than the total: any lens scored 1 rules a candidate out, and value must score at least 3. A safe, easy task that nobody cares about is a poor first automation, however well it scores elsewhere.
Shape the mission
Take the top candidate and shape it into a mission before committing:
- One workflow and one outcome, written in a sentence.
- What the agent may do, and what it must never do.
- Where a person approves, and who that person is.
- The baseline: volume, time, errors and cost today.
- The starting level of autonomy. Usually observe or assist: the agent recommends or prepares, and a person acts or approves. It moves up only on evidence.
If the mission cannot be written on one page, it is too big. Cut it until it can.
Decide, and record it
Make the decision explicit and record it: the chosen mission, the owner, the baseline, the target, the date on which results will be reviewed, and what result would mean expand, adjust or stop. The runner-up candidates are not discarded; they become the backlog for the second mission.
Before the agent goes live, test it against the mission with Evaluation Kit 001.
Good and poor first automations
Usually good:
- preparing and routing incoming requests for a person to handle;
- matching documents to records (invoices to orders, forms to accounts) and flagging mismatches;
- gathering context from several systems into one brief before a decision;
- drafting routine responses for a person to approve.
Usually poor:
- anything that commits money, prices or contracts without review;
- processes that change every month;
- tasks so rare that nobody will notice the improvement;
- "automate the whole department".
Traps when scoring
- Scoring without the people who do the work. Managers often overestimate how clear the rules are. Ask the people who handle the exceptions.
- The sponsor's favourite. A senior person's pet project scores well because nobody wants to mark it down. Score anonymously first, then discuss.
- Feasibility on a vendor's word. "The platform can do that" is not evidence. Feasibility means your data, your systems and your exceptions.
- Risk in the abstract. Score the specific actions the automation would take — reading, drafting, updating a record, sending, paying — not "AI risk" in general.
- Skipping readiness. It is the lens people find most uncomfortable to discuss, and often the one that decides whether the result is adopted.
Example
Illustrative example, not a client engagement.
A wholesaler screens its long list down to three candidates.
| Candidate | Value | Feasibility | Risk | Readiness | Result |
|---|---|---|---|---|---|
| Match supplier invoices to orders and deliveries | 4 | 4 | 3 | 4 | Chosen |
| Decide customer refund requests | 4 | 3 | 1 | 3 | Ruled out: risk scored 1 |
| Prepare the weekly sales report | 2 | 5 | 5 | 4 | Ruled out: value below 3 |
The sales report has the highest total, but nobody's week would change if it were automated. Refund decisions matter, but the value depends on an agent deciding about customers' money. Invoice matching has volume, clear rules, accessible data and a finance lead who wants it fixed. The agent starts by preparing matches for review; mismatches go to the buyer who raised the order.
Checklist
- The long list came from real queues, complaints and the people doing the work.
- Candidates without an owner, a measure or a designed process were screened out.
- Each candidate was scored on value, feasibility, risk and readiness.
- No chosen candidate scores 1 on any lens, and value is at least 3.
- The mission fits on one page, with what the agent must never do.
- Approval points and approvers are named.
- The baseline is measured, and the review date is set.
- The runners-up are kept as the backlog for the next mission.
