Published by the SELLIGNITE team. A practical operating model with examples and checks that show whether it is working. Examples are illustrative unless stated otherwise.
To choose your first AI use case, list the repetitive jobs your team does every week, then score each one from 1 to 5 on volume, time spent, clarity of the right answer, data access and cost of a mistake. Drop any job without a named owner or where one error could seriously hurt a customer. Pilot the highest scorer in review mode, with a person approving the output, before giving the agent more autonomy.
- A good first AI use case is frequent, time-consuming, easy to check and cheap to get wrong.
- Score candidates on five criteria from 1 to 5, for a maximum of 25 points.
- Two gates override the score: a named owner must exist, and a single error must not seriously harm a customer.
- Prefer internal, reversible work such as triage, extraction and drafts over autonomous customer-facing bots.
- Measure the pilot against a baseline: hours spent, turnaround time and the share of outputs approved without edits.
- 01List repetitive jobs
- 02Score five criteria
- 03Apply two gates
- 04Pick one pilot
- 05Review weekly
What makes a good first AI use case?
A good first AI use case is one specific, repetitive job that happens every week, takes real time, has an answer a person can check quickly and causes little damage when the output is wrong. It is not a department, a strategy or a tool. It is a job with an input, an output and someone who owns the result.
Small and mid-sized B2B companies usually get this choice wrong in one of two ways. Some start too big: "AI for customer service" or "an AI sales team", which nobody can scope, test or measure. Others start with whatever a vendor demo made look impressive, such as a chatbot on the website, which is customer-facing from day one and hard to supervise.
The first project matters more than its size suggests. It sets how your team feels about AI agents, it tests whether your data is usable, and it teaches you how to review an agent's work. A small, boring win that runs every day for months is worth more than an ambitious pilot that is switched off after three weeks. For the sales side of this question, our article on what an AI SDR can and cannot do shows which sales development tasks suit an agent and which do not.
How to find AI use case candidates in your company
Find candidates by looking for work people repeat, copy or wait for, not by asking where AI could be used. The question "where could we use AI?" produces wish lists. The question "what did you do three times last week that felt like copying?" produces real jobs.
Spend one week collecting candidates. Ask each team lead to write down every task that matches one of these signals:
- Copying between tools. Someone moves data from email into the CRM, from a PDF into the ERP or from a spreadsheet into a report.
- Waiting. Leads, replies, orders or tickets sit for hours before anyone looks at them.
- Repeated answers. The same customer or internal questions are answered from scratch every week.
- Sorting. A shared inbox such as sales@ or info@ is read by a person whose main job is deciding who should handle each message.
- Preparation. Reps research accounts or assemble meeting notes by hand before every call.
- Reporting. A weekly or monthly report is rebuilt manually from the same sources.
Write each candidate as one line in the form "When [trigger], [role] does [job] using [inputs], producing [output]." For example: "When an order email arrives, the sales assistant retypes the order lines into the ERP and replies with a confirmation." If you cannot fill in all four parts, the candidate is not yet a job; it is an idea. Aim for 8 to 15 candidates before scoring.
The AI use case scoring framework: five criteria
Score each candidate from 1 to 5 on five criteria: volume, time, clarity, data access and error cost. The first two measure value, the next two measure feasibility, and the last measures risk. A perfect score is 25.
| Criterion | Question to ask | Scores 1 | Scores 5 |
|---|---|---|---|
| Volume | How often does this job happen? | A few times a month | Dozens of times a day |
| Time | How many hours a week does it take in total? | Under 1 hour | More than 10 hours |
| Clarity | Can a person tell quickly whether the output is right? | Experts disagree on the answer | A clear right answer, written rules or past examples exist |
| Data access | Are the inputs digital and reachable by an agent? | Paper, phone calls or someone's memory | Email, CRM, documents or an API the agent can read |
| Error cost | What happens if the output is wrong? | Legal, financial or customer damage that is hard to undo | Caught in review, cheap to fix |
Score as a group of two or three people who know the work, not alone. Disagreement is useful: if one person gives clarity a 5 and another a 2, the job probably has unwritten rules, and those need to be written down before any agent can follow them.
Notice what the framework does not reward: novelty. Inbox triage, document data extraction and reply classification score well because they are frequent, checkable and reversible. That is exactly why they make good first projects, even if they are not exciting.
Two gates that override the score
Before you rank the scores, remove every candidate that fails either gate: no named owner, or an error that could seriously harm a customer. A high score does not rescue a job that fails a gate.
Gate 1: a named owner
Someone must own the agent's results: review its output, collect the cases it got wrong and update its instructions every week. This is usually the person who does the job today. If nobody can give the agent 30 to 60 minutes a week in the first month, the pilot will drift and be abandoned, however good the technology is.
Gate 2: no serious single-error harm
If one wrong output could send an incorrect price to a customer, approve a payment, reject a job applicant or give legal advice, it is not a first use case. Those jobs may suit an agent later, with strict review, but they are the wrong place to learn. An error cost score of 1 automatically fails this gate.
Data quality is a soft gate. If the job depends on CRM records that are out of date, the agent will repeat the mess faster. In that case, fix the fields first; our guide to CRM hygiene for outbound lists the fields and stages that matter.
How to read the score and pick one AI use case
Pick the highest-scoring candidate that passes both gates, and start with exactly one. Use these bands as a guide, not as precise science:
- 20 to 25: a strong first candidate. Scope it and pilot it.
- 14 to 19: promising, but one criterion is weak. Find the low score and ask whether it can be fixed cheaply, for example by writing down the rules (clarity) or connecting a mailbox (data access).
- Under 14: park it. Revisit after the first pilot, when you know more about what an agent can handle in your stack.
When two candidates tie, use three tie-breakers in order. Prefer internal work over customer-facing work, because mistakes stay inside the team. Prefer a job with an existing written process, because the agent has rules to follow from day one. Prefer the job whose owner is most interested, because their weekly review is what makes the pilot improve.
Resist the temptation to start two or three pilots at once. Each one needs instructions, review and measurement, and a small team rarely has the attention for more than one at the start. The second use case goes much faster once the first is running.
How to run the first AI pilot in two weeks
Run the first pilot in review mode: the agent does the job, a person approves every output, and you compare the results against a baseline you measured before you started. Two weeks is enough to see whether the job works, not to perfect it.
- Measure the baseline. Before the agent touches anything, record how many items arrive per week, how long the job takes and how long items wait.
- Write the job description. Trigger, inputs, allowed sources, output format, what the agent must never do and when it must hand a case to a person.
- Collect 20 to 30 real examples. Past emails, documents or records with the correct output. These become both instructions and a test set.
- Run in review mode. Every output goes to the owner for approval. Log each edit and each rejection with a one-line reason.
- Review weekly. Turn the most common rejection reasons into new instructions, then decide whether the job is ready for partial autonomy on the easy cases.
This is the same rhythm described in our guide to AI account research before outreach, where a person spot-checks briefs and rejection reasons become new rules. For a wider view of pilots in small sales teams, see where AI agents actually fit into a small sales team. If you want help scoping the pilot, the AI process automation and AI agents services both start with a free AI Audit call and take about two weeks to implement, followed by ongoing care and weekly iteration.
Example: scoring five candidates at a 40-person distributor
Imagine a B2B distributor with 40 people, a small sales team and a busy shared inbox. After a week of collecting jobs, the team shortlists five candidates and scores them together. The scores below are illustrative.
| Candidate | Volume | Time | Clarity | Data | Error cost | Total | Gates |
|---|---|---|---|---|---|---|---|
| Order emails retyped into the ERP | 5 | 5 | 4 | 4 | 3 | 21 | Pass |
| Sorting the info@ inbox | 5 | 3 | 4 | 5 | 4 | 21 | Pass |
| Weekly sales report | 2 | 2 | 4 | 3 | 4 | 15 | Pass |
| Website chatbot quoting prices | 3 | 2 | 2 | 3 | 1 | 11 | Fails gate 2 |
| Reviewing supplier contracts | 1 | 2 | 2 | 4 | 1 | 10 | Fails gate 2 |
Two candidates tie at 21. The tie-breakers decide: inbox sorting is internal, while order entry touches what customers receive, so mistakes in order lines are more expensive. The team also has a written routing list for info@, and the office manager who reads the inbox today wants to own the pilot.
So the first use case is inbox triage: the agent reads each message, labels it as order, quote request, support, supplier or other, drafts a routing note and assigns it to the right person, with the office manager approving every label for the first two weeks. Order entry becomes the second use case, and it starts faster because the inbox is already connected.
The chatbot quoting prices scored poorly and failed a gate, not because AI cannot answer website questions, but because unreviewed prices sent to customers are a bad place to learn. It can return to the list later, as an assistant that drafts answers for a person to send.
Template: a one-page AI use case card
Fill in one card per shortlisted candidate. If a field is hard to fill, that is usually the weak point of the job.
Job: When [trigger], [role] does [job] using [inputs], producing [output].
Owner: name, and 30-60 minutes a week reserved for review.
Baseline: items per week, hours per week, typical wait time.
Scores: volume _, time _, clarity _, data access _, error cost _ = total _ / 25.
Gates: named owner (yes or no), serious single-error harm (yes or no).
Never: what the agent must never do or send.
Hand-off: when the agent stops and passes the case to a person.
Success after two weeks: share of outputs approved without edits, hours saved, wait time.
Keep the cards. The ones you park today become the backlog for your second and third AI use case, and the scores are easier to update once you have one agent running.
Your implementation checklist
- Collect 8-15 repetitive jobs in one week
- Write each as trigger, role, job, inputs and output
- Score volume, time, clarity, data access and error cost
- Remove candidates without an owner or with serious error harm
- Pick one pilot and measure its baseline
- Run two weeks in review mode and log every rejection
What to measure: Compare the share of agent outputs approved without edits, hours spent on the job and item wait time against the baseline recorded before the pilot.
Frequently asked questions
What is the best first AI project for a small business?
The best first AI project is one repetitive internal job that happens every week, has a checkable right answer and is cheap to get wrong, such as sorting a shared inbox, extracting data from order emails or classifying replies. Customer-facing bots and decisions about money or people are better as later projects.
How many AI use cases should we start with?
Start with one. Each AI use case needs instructions, an owner who reviews the output and a baseline to measure against, and a small team rarely has attention for more than one pilot at once. The second use case usually goes faster because the first one has already connected your tools and taught the team how to review an agent.
How long does a first AI pilot take?
A focused first pilot can be set up and running in review mode in about two weeks, if the inputs are digital and the owner has time to review. After that, expect several weeks of weekly iteration in which rejection reasons are turned into new instructions before the agent handles easy cases on its own.
Do we need clean data before starting with AI?
You need clean data for the specific job you pick, not for the whole company. If the job reads CRM records, the fields it uses must be current; if it reads emails or documents, those are often usable as they are. Scoring data access honestly shows where cleanup has to come first.
What if no candidate scores above 20?
Look at the lowest criterion of your top two candidates and ask whether it can be fixed cheaply, for example by writing down the rules or connecting a mailbox. If none can, run a smaller pilot on the best candidate anyway, in strict review mode, and treat it as a learning project rather than a time-saving one.