← All essays

Forecasting when AI touches every deal

Rebuilding the forecast model so the number survives contact with an AI-assisted sales motion.

The forecast you trusted last year is lying to you now, because AI has changed what every signal in your CRM means. When a model writes the call notes, fires the first touch, and logs the next step, the data still flows in clean. It just no longer means what it used to. Your job is to rebuild the model so the number survives contact with a sales motion where AI touches every deal.

This is a methodology problem, not a tooling problem. The coverage ratios, velocity, and conversion math still apply (I work that math in a separate piece). What breaks is the trust layer underneath the math: the assumption that activity reflects human intent and that a stage means what the playbook says it means. Once AI inflates the top of the funnel and ghostwrites the record, you have to re-earn that trust signal by signal.

The old signals lie now. Name how.

Start by being honest about what AI has quietly corrupted. Not maliciously. Just mechanically.

Activity volume used to be a rough proxy for effort and interest. A rep who sent 40 touches and booked 12 meetings was working a live patch of pipeline. Now a model can generate hundreds of personalized touches before lunch, and meeting-booking rates on AI-sourced outreach can look healthy while intent runs thin. The volume is real. The signal it used to carry is gone.

CRM notes used to encode judgment. When a rep wrote "champion confirmed budget, legal is the only risk," that sentence cost them something to write and reflected what they believed. Now an AI summarizer turns a 30-minute call into a tidy paragraph that reads like progress whether or not progress happened. The notes got better. Their value as a tell got worse.

Stage progression used to require friction. Moving a deal to "proposal" meant a human decided it belonged there. Now suggested stage updates, auto-logged next steps, and AI-drafted recaps make it cheap to look advanced. Cheap signals inflate.

If you do not name these three corruptions out loud with your team, everyone keeps reading the dashboard as if it still tells the truth. It does not.

Separate AI-generated activity from human intent

The first rebuild is structural: tag the source of every signal so you can tell machine motion from human motion.

You want to know, for any deal, what the AI did and what a human (rep or buyer) actually did in response. Most teams have the data and never split it. Build the split.

• Outbound: separate AI-generated touches from human-personalized follow-up. Track reply and meeting rates for each. AI gets you reach; humans convert. If a cohort is all reach and no human response, it is not pipeline, it is noise wearing a pipeline costume.
• Notes and recaps: mark which records are AI-summarized versus rep-authored on the fields that matter (next step, risk, decision criteria). An AI recap is fine as a transcript. It is not evidence the rep understands the deal.
• Buyer-side action: this is the gold. Did the buyer reply, forward, open the proposal, bring a second name to the table, propose a date. Buyer effort is the one thing AI on your side cannot fake.

The principle is simple. Your AI can manufacture activity on your side of the table all day. It cannot manufacture engagement on the buyer's side. Re-weight your forecast toward the side of the table you do not control.

Re-baseline conversion. Your historical rates are stale.

When the motion changes, the conversion rates you forecast against stop being valid. You cannot apply last year's stage-to-stage math to this year's AI-assisted funnel and expect the number to hold.

Here is the trap. AI usually pushes more volume into early stages while late-stage human-driven conversion stays roughly the same or even dips (more marginal deals got in the door). If you keep applying old conversion rates to a fatter, lower-quality top, the model over-forecasts. The pipeline looks bigger and converts worse, and the gap shows up as a miss at the end of the quarter.

Re-baseline deliberately:

1. Pick a clean cohort from after the motion changed. Do not blend pre-AI and post-AI deals into one average. You will get a number that describes neither world.
2. Recompute stage-to-stage conversion on that cohort. Expect early-stage rates to drop. That is the AI-volume effect, not a rep problem.
3. Segment by source. AI-sourced, human-sourced, inbound, and partner deals now convert differently enough that one blended rate hides the truth. Forecast each lane on its own rate.
4. Re-baseline quarterly until it stabilizes. The motion is still moving. A rate set in stone for a year will be wrong by month three.

This is unglamorous work. It is also the single highest-leverage thing you can do, because every downstream forecast number inherits these rates. Get them wrong and everything built on top inherits the error.

Pick leading indicators that survive an AI first touch

You need indicators that still mean something when a model made the first move. The test for a good indicator now: could AI on our side fake it. If yes, demote it. If it requires the buyer to spend effort or a human to exercise judgment, promote it.

Indicators that hold up:

• Multithreading. Count of engaged contacts on the buyer side, not contacts in the CRM. AI can populate a contact list. It cannot make three people from the buyer's org show up and ask questions. Single-threaded deals in an AI-heavy funnel are more dangerous than they used to be, because the top got easier to fill and the depth did not.
• Verified next step. A next step is only real if it exists on the buyer's calendar with a date and the buyer's agreement. "Following up next week" auto-logged by an assistant is not a next step. A booked meeting the buyer accepted is.
• Buyer-side engagement. Proposal opens, doc shares forwarded internally, security or legal review initiated, pricing questions from a finance contact. Effort the buyer spends is the cleanest intent signal left.
• Champion-tested risk. Did a human rep name the specific risk and the path to clear it. Not the AI recap of the risk. The rep's own read. This is judgment, and judgment still predicts outcomes.

Indicators to demote: raw activity counts, email opens, AI-summarized "positive sentiment," and stage age on its own. They were always weak. AI made them weaker and easier to game.

Tighten category discipline so the model means something

Commit, best case, and pipeline only work as forecast categories if they carry strict, enforced definitions. AI-inflated pipeline makes loose definitions fatal, because there is now far more low-intent volume sitting in the funnel waiting to be miscategorized as something better than it is.

Set the gates and hold them:

• Commit: verified next step, multithreaded, mutual close plan with buyer-confirmed dates, and a named human owner of every open risk. If any of those is missing, it is not commit. No exceptions because the AI recap "sounded good."
• Best case: real buyer-side engagement and a credible path, but at least one of the commit gates not yet cleared. Upside, not plan.
• Pipeline: everything else, including the entire pile of AI-sourced deals that have not yet drawn a single human response from the buyer. Most of it will not close. Forecast it accordingly.

The discipline that matters most: a category is a claim about the buyer, not about your team's activity. AI can make your side look busy in any category. The gate is always what the buyer did.

Run the forecast call so reps cannot hide behind AI pipeline

The forecast call is where the rebuilt model gets enforced or quietly abandoned. In an AI-assisted motion, the failure mode is reps pointing at a fat, model-generated pipeline as evidence the quarter is fine. Your call has to make that impossible.

Change the questions. Old forecast calls asked "what's the status." That invites a recap, and AI writes great recaps. Ask instead for the buyer-side evidence:

• Who on the buyer side took an action this week, and what was it. Not what we sent. What they did.
• What is the next step, who booked it, and is it on their calendar. Show me.
• Who else from their org is engaged. Name them. If it is one person, why is this a commit.
• What is the single biggest risk, and what is your plan to clear it. In your words, not the recap.

Make reps cite buyer-side proof for every commit deal. The standard is evidence, not narrative. A rep who can only produce AI-generated activity and a clean-looking summary does not have a commit. They have hope with good formatting.

The forecast call is no longer a status meeting. It is an evidence review. If the only proof a rep can show was generated by your own systems, the deal does not count.

Inspect the source split live. When a rep calls a deal, pull up whether the recent activity was AI-generated or human, and whether the buyer responded. The first few times you do this, the room gets quiet, because everyone realizes the dashboard they leaned on was measuring their own machine. That quiet is the point. It resets what counts as real.

The takeaway

AI did not break forecasting. It broke the proxies forecasting quietly relied on, and it did so without changing how the data looks. The fix is not a better tool. It is a disciplined rebuild: split machine activity from human intent, re-baseline conversion on a clean post-change cohort, promote indicators the buyer has to earn, enforce categories as claims about the buyer, and run a forecast call that demands evidence over recap.

Do this and the number gets harder to produce and far more honest. In a motion where a model touches every deal, the leaders who win the forecast are the ones who stop trusting activity and start demanding proof of the one thing AI on their side can never manufacture: a buyer who is actually moving.