Start with a task your team repeats often, whose inputs and expected result are clear, and where mistakes can be caught before they cause harm. Look for repeated data entry, routine handoffs and time spent finding the same information in different systems.
Before choosing a tool, check whether the task needs to happen at all. A shared record or a simpler approval process may remove work that you were about to automate.
Your first project should be small enough to test, useful enough for staff to adopt and measurable enough to tell whether it helped. Here is a practical way to choose it.
1. Find the work that keeps coming back
Ask the people doing the work to describe three tasks they repeat. Follow each task from its trigger to its finished result: what arrives, what someone does with it, and who needs the output.
Look for:
- Information copied from one application into another.
- Repeated searches for a customer, order or project record.
- Documents assembled from information you already hold.
- Handoffs that depend on someone remembering to send an update.
- Routine checks with a clear rule for what happens next.
Write down a specific task, such as “enter a new programme into the back-office and website catalogues”, rather than a whole department's activity, such as “automate sales”. You need a boundary you can observe and test.
Record how often the task happens and how much active staff time it takes. Keep waiting time separate: a request sitting in an inbox for two days is a different problem from a request that takes two hours to process.
2. Simplify the process before automating it
Repeated work can be a sign that information is stored in the wrong places. Connecting records may be more useful than making duplicate entry faster.
At Ampersand, the back office and website administration used separate applications and databases. Programme information was entered in both places and needed to be kept aligned. Buyer viewing accounts and client catalogues lived in the website administration system, away from the deals and relationships managed in the back office.
The new CRM brings those records together. A programme has one record shared by sales and the public website. A buyer's viewing account and programme selections sit alongside their company information and deals. Atea is the lead consultant on the project.
This is a useful example of the question to ask first: can we remove the reason for the repeated work? Keeping two copies synchronised is one possible approach. Sharing a record removes the need to maintain the second copy.
The Ampersand CRM is built and tested but is not yet in use; migration and first staff deployment are being prepared. These are intended workflow improvements, not measured results. Read the Ampersand case study for the project detail.
For your own process, ask which information needs one agreed source, which approvals still serve a purpose, and which steps exist only because two systems do not connect.
3. Compare candidates with a simple worksheet
Use the worksheet below for each task. Record what you know before scoring it; mark unknowns rather than guessing. The scoring is a discussion aid, not a validated assessment or a return-on-investment forecast.
| Factor | What to record | 0 points | 1 point | 2 points |
|---|---|---|---|---|
| Frequency | Occurrences in a typical week or month, including peaks | Occasional or unpredictable | Regular, usually weekly | Usually daily or several times a day |
| Handling time | Occurrences × active minutes per task, including checks and rework | Little total time relative to other candidates | Some recurring time | One of the larger recurring demands on the team |
| Consistency | How often the same steps lead to an agreed result | Each case needs a different approach | A repeatable core with known variations | Clear, stable steps and result |
| Data readiness | Where inputs live, who owns them and whether they are complete | Inputs unclear or unreliable | Some cleanup or access work needed | Accessible, consistent inputs with an agreed source |
| Exceptions | Cases that need intervention and who handles them | Exceptions are hard to identify or assign | Known exceptions need frequent attention | Exceptions are identifiable and have a clear human route |
Total: ___ / 10. Compare tasks within your business, using the same period and assumptions. A higher score suggests a candidate worth examining; it does not approve a project.
Complete these checks separately. A high score cannot cancel out a serious consequence of error.
| Decision check | Your notes |
|---|---|
| What happens if the output is wrong? Consider customers, money, confidential information and recovery work. | ___ |
| Can someone check the output before it takes effect? | ___ |
| Who owns the process and handles failures? | ___ |
| What access, integration or data cleanup is needed? | ___ |
| Could a simpler process or an existing tool remove the task? | ___ |
| What would make the pilot worth continuing, and what would make you stop it? | ___ |
For example, duplicate entry across two catalogues is worth investigating because its trigger and records can be identified. You would still need to measure the volume and handling time, agree which record is authoritative, and check how exceptions work before choosing an implementation. That is the reasoning illustrated by Ampersand; it is not a retrospective score for the project.
4. Decide whether the task needs AI
If the work follows clear rules—copy an approved field, update a status or generate a document from agreed data—ordinary software or an integration may be enough.
AI may be useful when the task involves interpreting variable text or preparing an analysis for someone to review. That introduces another question: how will a person check whether the output is useful and correct?
Ampersand's separate scriptwriting scenario application illustrates this distinction. Writers can propose changes, and Claude analyses their impact on the rest of the story. The writers retain approval. This application is separate from the CRM and is also built and tested but not yet in use.
An AI pilot needs to include the time spent reviewing and correcting its output. A quickly generated answer is not a useful saving if checking it takes longer than the original task. If it provides analysis that was not previously done, assess that added capability separately from time saved.
5. Choose one small pilot
Define one start point, one intended result and a limited set of cases. Give the pilot an owner and an agreed review point.
Before it starts, write down:
- The task and which cases are included or excluded.
- The baseline: current volume, active handling time, errors and rework.
- The checks required before an output changes a record or reaches a customer.
- How staff recognise exceptions, pause the process and complete the work manually.
- The evidence you will use to continue, revise or stop the pilot.
Use representative work, including awkward cases. A demonstration on one clean example does not show how the process will behave in a busy week.
Keep the first pilot's scope narrow enough that staff can explain what changed and tell you where it fails. Expand only when that evidence supports the next step.
Example: preparing information for an accountant
A proposed Tari Digital project for small UK accounting practices shows how to narrow an ambitious idea. The aim is to help clients complete their accountant's existing Self Assessment checklist, provide supporting documents and resolve missing information, so the accountant receives an organised preparation pack.
The first version would work without AI: an online checklist, secure document uploads, completion tracking and an accountant review screen. That would establish the workflow before adding document interpretation.
A later pilot would test extracting figures from documents, linking each value to its source document and page, and flagging missing or inconsistent information. Conventional PDF extraction and OCR would provide the baseline. Locally hosted AI would be tested where interpretation is needed, and retained only if it adds measurable value after checking and correction time are included.
The proposed boundary is explicit: confidential client documents must not be sent to external LLM services, and the accountant must verify and approve every figure before finalisation. Access controls, storage, retention and hosting arrangements would need review before real client records are introduced; self-hosting alone would not establish that the application is secure.
Initial tests would use synthetic documents. The pilot would compare extraction accuracy, correct source references, processing time and the accountant's review effort. It would prepare information for review; direct HMRC submission, autonomous tax calculations and tax advice are outside the proposed scope.
This is a project concept, with no completed implementation or measured savings claimed. Its lesson for a first automation project is practical: choose one administrative stage, establish the checks, then test whether AI improves that stage.
6. Measure the whole task
Compare similar work before and during the pilot. Count checking, corrections and exception handling as part of the new process. Record training and setup separately so you can distinguish the initial effort from ongoing use.
A simple measure is:
Capacity released = (baseline active minutes per task − pilot active minutes per task) × comparable task volume.
Convert minutes to hours by dividing by 60. Keep task boundaries consistent and avoid counting the same minutes twice—for example, under both “less searching” and “faster updates”.
Also check whether staff use the process, whether errors have changed and whether customers or colleagues receive a better result. Time released is capacity the team can use elsewhere; it is not automatically a reduction in cash costs. Setup, maintenance and running costs still belong in the decision to continue.
Start with one workflow you can explain
The best first candidate is a recurring task with understandable inputs, an agreed result and a manageable way to handle mistakes. Simplify it, test a small change and measure the complete job.
Tari Digital helps owner-led businesses assess AI and workflow automation. If you have a task in mind, bring a short description of what starts it, where information gets entered twice and who checks the result.
Request a free workflow review.
The free review is an introductory conversation about one repetitive task and a practical next step. Detailed process mapping, technical investigation and a written implementation plan are part of a separately agreed paid assessment. Implementation is scoped separately.