Fix the record before you add agents

The pilot looked great in the demo. An AI agent would read every inbound lead, research the company, score it and route it to the right rep within minutes. Two weeks after go-live, the sales team asked for it to be switched off.

It had routed a 2,000-person prospect to the SMB team because the employee count field was empty. It sent a follow-up to a contact who had left the company a year ago. It assigned three leads from the same account to three different reps, because the account existed three times in the CRM. The agent did exactly what it was told. The record it was reading was wrong.

This is the most common way AI projects in revenue teams fail, and it has very little to do with AI.

Why the foundation matters more now

A rep working from a bad record can use judgement. They notice the duplicate, they know the contact left, they ask a colleague. An agent cannot. It treats the CRM as the truth, and acts on it quickly, confidently and at volume.

That changes the economics of data quality. When people did all the work, a messy CRM cost you some time and some trust in reports. When agents do the work, a messy CRM produces wrong actions: the wrong lead routed, the wrong person emailed, the wrong history summarised for the forecast call. The cost of a bad record goes up with every workflow you automate.

The symptoms are usually visible long before anyone mentions AI:

• Reports that do not reconcile. Sales, marketing and finance each have a different pipeline number, and every board meeting starts with a debate about which is right.
• Reps keeping their own spreadsheets. When people stop trusting the CRM, they track the real state of deals somewhere else, and the record falls further behind.
• Leads that wait. Inbound leads sit unassigned for hours or days because the routing rules do not cover the case, or the account owner field is blank.
• Manual fixes before every forecast. Someone spends a day each month correcting stages and close dates so the numbers look sensible.

Each of these is a person quietly compensating for a bad record. An agent cannot compensate. It inherits every one of these problems and repeats them faster.

The good news is that you do not need a perfect CRM. You need the specific data that one workflow depends on to be right, owned and kept right.

The five checks to run first

Before any agent touches your CRM, run these five checks. Each has a simple test you can do this week.

1. One record per account and person. Duplicates are the fastest way to break routing, scoring and reporting. Test: export accounts, normalise the website domain, and count domains that appear more than once. Above 5% is a problem.

2. Fields that mean one thing. Most CRMs have hundreds of custom fields and a dozen that actually drive decisions. Test: list the fields your routing, scoring and forecast depend on. For each, can two people write the same definition? If not, an agent cannot use it.

3. Stages with exit criteria. "Proposal" should mean the same thing in every team, with a clear rule for moving in and out, signed off by sales and finance. Test: pick ten open deals in the same stage and ask whether each one meets the written criteria.

4. A source and date on every enriched value. Employee count, industry and revenue come from somewhere, and go stale. Test: for your key firmographic fields, can you tell where the value came from and when it was last checked?

5. Owners and handoffs written down. Routing needs rules: who owns which territory, segment and stage, and how fast they must respond. Test: ask three managers what happens to an inbound lead from a 500-person company in Germany. If you get three answers, your agent will too.

Before and after: what the agent sees

The difference between a failed pilot and a working one is rarely the model. It is what the model reads.

How to tell if a workflow is ready for an agent

Not every workflow needs the same foundations. Meeting prep can tolerate a few stale fields; lead routing cannot. Before you automate anything, walk it through four questions.

Fix one workflow at a time

The mistake most teams make is turning data quality into a programme. A twelve-month "CRM clean-up" with no visible result loses sponsorship by month three. Scope it the other way round: pick the workflow, fix only what it needs, prove the result, then move on.

Step 1: pick the workflow

Owner: CRO and Head of RevOps. Done when: one workflow is named, with a number you want to move.

Good first candidates are high volume and easy to measure: inbound lead routing (speed to first touch), meeting prep (rep hours saved), or pipeline hygiene (deals with a next step). Avoid starting with forecasting; it depends on everything else.

Step 2: map the data it reads

Owner: RevOps. Done when: a list of every field, object and rule the workflow touches.

For inbound routing that is usually around 10 to 15 fields: company domain, employee count, country, industry, lead source, existing account match, account owner, territory, segment and the round-robin rules. Write each one down with its definition.

Step 3: fix only that data

Owner: RevOps, with sales managers for the rules. Done when: those fields pass the five checks above.

Merge duplicates on the matched accounts. Backfill the firmographics from one trusted source and store the date. Write the routing rules as a table, not a paragraph. Add validation so new records cannot break them.

Step 4: prove it, then add the agent

Owner: the workflow owner. Done when: four weeks of results against the baseline.

Run the workflow with the clean data first, and only then add the agent with a person reviewing its output. Measure speed, accuracy and how often the reviewer changes the result. When the edit rate is low and stable, widen what the agent can do on its own.

How long it takes

For most mid-market teams, the first loop takes weeks, not quarters, when it is scoped to one workflow. The second is faster, because much of the data overlaps: the account and contact clean-up you did for routing also serves meeting prep and renewals.

The payoff shows up quickly. Reports reconcile because stages mean one thing. Reps trust the CRM again because the record matches reality. And every agent you add afterwards starts from solid ground instead of inheriting years of shortcuts.

Where AI helps with the clean-up itself

There is a useful twist here: AI is good at the foundations work too, as long as a person approves the changes. An agent can find likely duplicates, suggest which record to keep, fill missing firmographics from public sources with the source attached, and flag deals that do not meet their stage criteria. What it should not do is merge records or rewrite stages on its own. Clean-up changes are hard to undo, so they go through review.

A practical set-up is a weekly queue: the agent proposes merges, backfills and stage corrections with the evidence for each, and RevOps approves them in batches. Track how many proposals are accepted unchanged. Once that acceptance rate is high and stable for a type of change, such as filling a missing country from a verified source, it can be applied automatically while riskier changes stay in the queue.

Keep it clean once it is clean

A one-off clean-up decays within months unless new data is held to the same standard. Three habits keep the record right:

• Validation at the point of entry. Required fields, picklists instead of free text, and a duplicate check when someone creates an account.
• A named owner for each critical field. Someone who is told when it is blank or stale, and has the job of fixing it.
• A short monthly health check. Duplicate rate, blank critical fields and deals without a next step, reviewed in the same meeting as pipeline.

Where to start this week

If you would rather not do it alone, our RevOps foundations work covers exactly this: the record, the definitions and the routing rules for your first workflow, so the agents you add next have something solid to read.