How to stop pipeline inflation from AI-created deals

By KYN AI Advisory Team — AI implementation specialists, Singapore

An agent that creates deals on its own can inflate your pipeline with opportunities that are not real. The forecast grows, and the win rate quietly falls. Three controls prevent it: require two independent signals before a lead becomes a deal, log where every deal came from, and let a named person confirm anything the agent is unsure about. A missed deal costs one lead. A false deal distorts every forecast that includes it.

How does an agent inflate a pipeline?

An agent that reads email, forms and WhatsApp can open a deal each time it sees interest. Many of those signals are weak: an out-of-office reply, a question from a student, a forwarded newsletter, a vendor pitching you. Each becomes a record with a stage and a value.

Individual records look harmless. In aggregate they raise pipeline coverage, push stage-conversion rates down, and send reps after contacts who never intended to buy.

Why is a false deal worse than a missed one?

Reconciliation work gives the rule. A false match is worse than an open item, because it looks resolved and takes longer to detect.

| Error | Visible cost | Hidden cost | |---|---|---| | Missed real deal | One lost opportunity, noticed on review | Bounded | | False deal created | A rep's time | Forecast error, wrong hiring and cash plans, lost trust in CRM data |

Bias the gate toward holding a lead rather than promoting it.

What is a two-signal gate?

A lead becomes a deal only when two independent signals agree. Pick signals that do not share a cause.

| Signal type | Examples | |---|---| | Intent | A reply that asks about price, scope or timing; a meeting request; a quote request | | Fit | Company in your target size and industry; a role with buying authority; a registered business | | Context | An existing relationship; a referral; a prior deal at the same account |

Examples of the rule:

• Intent plus fit creates a deal at an early stage.
• Intent alone creates a lead flagged "unverified", not a deal.
• Fit alone creates nothing.

Widen the rule only with an independent check. A different job title may be acceptable if a prior deal at the same account confirms it. Never widen on tolerance alone.

What should the agent record?

Every agent-created deal needs provenance, written at creation.

• Source message or event, with a link.
• Which two signals fired, and their values.
• The agent account that acted, and the time.
• Confidence label and the rule version.
• Current owner, a named person.

With this, a forecast review can filter by "created by agent" and see which signals were weak. Without it, you can only argue.

How do you audit what is already in the pipeline?

Run an ask-over-assume review on agent-created deals.

1. Pull deals created by the agent in the last 90 days.
2. Sample 50. Have a person mark each real, unclear or false, with a reason.
3. Compute the false rate by source and by signal pair.
4. Tighten the gate where false rates cluster.
5. Repeat monthly.

If you cannot pull agent-created deals as a separate set, that is the first fix.

Which guardrails belong at the point of creation?

• A confidence label on each deal: confirmed, probable or unverified. Only confirmed and probable count toward the forecast.
• A cap on deals the agent may open per day, with an alert on a spike.
• Duplicate checks against existing accounts and contacts before creation.
• A review queue for unverified items, with a one-click confirm or reject.
• Deletions and merges wait for a named person.

Review stays quick when the reviewer sees the source message, the signals and a single decision. A batch clears in minutes.

What does clean master data have to do with it?

Everything. If accounts and contacts are duplicated or inconsistent, the agent opens a new deal on a company that already has one. Agree the account and contact structure first, add role-based views so sales sees its own pipeline and finance sees a forecast built from confirmed deals, and then switch on creation.

What should you look for in a deal-creation tool?

• Does it log provenance on each record?
• Does it write to your CRM in real time, not in a nightly batch?
• Can you change the gate rules without a developer?
• Does it separate agent-created deals in reporting?
• Does it hold an unsure lead instead of promoting it?

How much can false deals distort a forecast?

The numbers below are illustrative. They show the mechanism, not a result from any client.

A sales team has 40 genuine open deals averaging S$20,000, so S$800,000 of pipeline. It historically closes one deal in four at this stage, which gives an expected S$200,000. An agent is switched on and opens 30 extra deals a month from weak signals. After two months, 60 extra deals sit in the pipeline at the same S$20,000 average. None will close.

| | Before agent | After two months | |---|---|---| | Open deals | 40 | 100 | | Pipeline value | S$800,000 | S$2,000,000 | | Deals that will close | 10 | 10 | | Observed close rate | 1 in 4 | 1 in 10 | | Forecast at the old one-in-four rate | S$200,000 | S$500,000 |

The forecast rises by S$300,000 and nothing real has changed. A manager who reads the pipeline figure plans hiring and cash against revenue that will not arrive. A manager who reads the falling close rate concludes the sales team has got worse. Both readings are wrong, and both are caused by the same false records.

What do the five most common weak signals look like?

An agent reads messages, not intent. These five message types look like buying interest and often are not.

| Message | Why it looks real | What it usually is | Gate result | |---|---|---|---| | Out-of-office reply | Contains a company name and a date | An automatic reply | No deal | | "Please send your brochure" | A request | Early research, often a student or a competitor | Lead, unverified | | Reply from a vendor | Mentions "partnership" | A sales pitch aimed at you | No deal | | Forwarded newsletter | A contact engaged with your content | A passive read | Lead, unverified | | "Can we get a quote for 10 units?" from a target-fit company | Specific, bounded | Likely real | Deal, early stage |

The last row is the only one that satisfies both intent and fit. The gate encodes that difference.

How should the gate score a signal?

Keep it explicit. A small points table is easier to defend than a model score nobody can inspect. The values below are illustrative; set your own from the audit.

| Signal | Points | |---|---| | Asks for price, scope or timing | 2 | | Requests a meeting | 2 | | Company matches target size and industry | 1 | | Contact has a buying role | 1 | | Existing relationship or referral | 1 | | Prior deal at the same account | 1 |

Rule: create a deal at 3 points or more, and require at least one point from intent and one from fit or context. Below 3, the agent creates an unverified lead. The rule version is recorded on every deal, so when you change the table you can compare the before and after.

Points are a convenience. The independence rule matters more. Two signals from the same message, such as "price" and "quote", are one signal.

What states should a deal pass through?

A deal that the agent opens is not yet a deal in the forecast. Give it a state.

1. Unverified lead. One signal. Not in the pipeline. Visible in the review queue.
2. Probable. Two independent signals. In the pipeline at an early stage, counted at a lower weight in the forecast.
3. Confirmed. A named person has spoken with the contact, or the contact has met the criteria you set, such as a meeting held. Counted in full.
4. Rejected. A person marked it false. The reason is stored, and the pattern feeds back into the gate.

Only confirmed and probable deals feed the forecast. The forecast report states the share that is probable, so a reader can see how much of the number is unconfirmed.

What does a good audit sheet contain?

A person labels each sampled deal. The columns below take about 90 seconds per row.

| Column | Content | |---|---| | Deal ID | Link to the record | | Source message | The text that triggered creation | | Signals fired | The two signals and their values | | Reviewer verdict | Real, unclear or false | | Reason | One line, from a short list | | Rule version | Which gate created it |

Reasons should come from a list: out-of-office, vendor pitch, student or research, duplicate account, wrong contact, test message, other. The list shows where to fix the gate. If "duplicate account" leads, fix master data. If "vendor pitch" leads, add a sender-domain check.

Read the result by signal pair, not just overall. An overall false rate of about one in seven hides the pair that is wrong half the time.

How should the pilot roll out?

1. Shadow mode, two weeks. The agent writes proposed deals to a review table. Nothing enters the CRM pipeline. A person labels a sample.
2. Limited creation, two weeks. The agent creates deals for one team and one source, with a daily cap. Review all of them.
3. Widen by source. Add one source at a time: forms first, then email, then WhatsApp. Each source has its own false rate.
4. Widen by team. Move to the next team after the false rate is stable for a month.

At each step, define the stop condition in advance. For example: if more than a set share of the sample is false, the agent returns to the previous step. You set the share. A condition written after the fact is not a condition.

What are the edge cases?

• Reopened closed deals. A contact who lost a deal replies months later. The agent should reopen the old deal, not create a second one.
• Multiple contacts, one company. Two people from the same company write in. Link them to one account and one deal.
• Referrals. A referral is strong context but not intent. Treat it as one signal and still require another.
• Existing customers. A support question from a customer is not a new deal. Route it to support. An expansion request is a new deal, and a person confirms it.
• Bulk outreach replies. After a campaign, replies spike. A daily cap and an alert on spikes stop the agent from opening dozens of records in an hour.
• Test data. Internal tests create records. Mark test accounts so the gate ignores them.

What should you report each month?

• Agent-created deals by state: unverified, probable, confirmed, rejected.
• False rate in the audit sample, by source and by signal pair.
• Time from creation to a person's confirmation.
• Share of the forecast that is probable.
• Close rate of agent-created deals against human-created deals, once enough have closed to compare.

The last figure is the honest test. If agent-created deals close at a much lower rate than human-created ones, the gate is too loose, whatever the sample says.

What access should the agent have?

Give the agent its own restricted account in the CRM. It may create leads and deals in the states above, add notes and read account data. It may not change a deal to won or lost, edit a value after a person has set it, merge accounts or delete anything. Each of those waits for a named person. Because every write carries the agent's account and a timestamp, you can filter the pipeline to see exactly what it touched.

Keep the gate rules in a configuration file or table that a named person can change, with version history. When a rule changes, record who changed it and why, so the audit sheet can compare rule versions fairly.

What is the trade-off between a tight and a loose gate?

A tight gate misses real deals. A loose gate inflates the pipeline. You cannot remove both errors, so you choose where to put the cost.

| Gate setting | What improves | What gets worse | |---|---|---| | Tight (three signals) | Clean forecast, little review work | More real leads wait as unverified; slower first contact | | Medium (two signals) | Balance of speed and accuracy | Needs a monthly audit to stay honest | | Loose (one signal) | Fast capture, every interest recorded | Inflated pipeline, wasted rep time, distrust of CRM data |

For most SMEs the medium setting is the right start, with the unverified queue as the safety net. A missed lead in the queue is still visible to a person. A false deal in the pipeline is invisible until the forecast misses.

What should be checked before go-live?

Use this list on the day before the agent first writes to the pipeline.

• Every agent-created deal can be filtered as a separate set.
• Each deal stores its source message, signals, rule version and time.
• The daily creation cap is set, with an alert on a spike.
• The duplicate check runs against accounts and contacts before creation.
• Deletion, merge, won and lost changes are blocked for the agent's account.
• A named person owns the review queue, with a deputy.
• The forecast report excludes unverified leads and labels probable deals.
• A written stop condition exists, and a person has the authority to switch the agent off.

If any line is false, delay the launch. Each is a small piece of work now and a large piece of cleanup later.

What if reps ignore the review queue?

A queue nobody reads is a queue that fails. Reps ignore it for two reasons: it has too many items, or the items are not worth their time. Fix the first by raising the gate. Fix the second by showing only what a person must decide, with the source message beside it. Assign one reviewer per day instead of asking everyone to check. Measure the age of the oldest item, and treat an item older than your service target as an incident. Report that age to the sales lead each week. A queue that grows by ten items a week and is cleared once a month produces the same distortion as having no gate, only delayed. If the volume stays high after the gate is tuned, add a second reviewer or lower the daily creation cap, and record the change with its date so the next audit can account for it.

Where KYN fits

KYN builds qualification agents on top of the CRM a client already uses. For an insurance brokerage, outreach runs from one checkable first line at about S$0.02 per email and replies are sorted at about S$0.01 each. Every action is logged with the actor. Agents have their own restricted accounts, and outbound messages wait for a named person.

The first 14 days are a review: an inventory of your stack and interviews with the team. A full build ships in under 90 days. See sales systems and custom CRM data models.

Frequently asked questions

Should an AI agent ever create a deal automatically?

It can, with two independent signals and a confidence label. Single-signal leads should be held for a person.

How many false deals are acceptable?

Set the threshold from your audit, not from a benchmark. Decide how much forecast error you can tolerate and tighten the gate until the sample meets it.

Does this slow the sales team down?

Slightly at creation, less afterward. Reps spend less time on contacts who never intended to buy.

Checked 30 October 2026.

Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.

Start Building →