Executive summary
Your CRM already holds the signals that predict revenue, but they're measured wrong, and AI is learning from the mistakes.
AI forecasting and deal-scoring tools learn from deal history: when each deal was expected to close, how often that date moved, which stages it reached and whether it won. The quality of that history sets the ceiling on every score and forecast those tools produce.
We tested six years of B2B CRM deal history against the questions those models depend on, and audited 52 public AI readiness checklists to see who else asks them. The history held more predictive signal than the team was using, and 96% of the checklists never look for it.
Six findings
- 01Commit lost more deals than it won.52%of deals called Commit were lost
- 02First forecasts almost never held.97%of wins missed their first-forecast month
- 03Most pipeline outlived the sales cycle.68%of pipeline open 1.5x past the average win
- 04The fields that predict wins sat empty.2.6xwin rate when an Economic Buyer is named
- 05Losses left a model nothing to learn.81%of lost deals gave no usable reason
- 06AI now edits history nobody tests.96%of readiness checklists never test it
52 checklists44 publishers6 years of deal history733,834 CRM records1,021,915 pipeline snapshots
About the data. An anonymized B2B CRM dataset spanning 2020 to 2026, analyzed by AeolusGTM. Results are shown as percentages and ratios, with sample sizes in the methodology. No dollar values are reported.
Forecast categories
Finding 01 / 06
Commit lost more deals than it won, and almost no manager weighed in.
52%
of deals called Commit were lost.Commit deals still won four times as often as the average deal.
Commit is the category leadership treats as safe, the deals a CRO puts in front of the board. Those deals did carry signal, winning four times as often as the average deal, yet more than half still closed lost. An AI forecast trained on this history learns that a Commit call is worth about half a deal.
The human check that should catch the gap was missing. The manager forecast field was blank or marked "Omitted" on 98% of open deals, so neither the team nor a model can see where a manager disagreed with a rep.
More than half of Commit deals and four in five Upside deals closed lost
Reached CommitEach square is 1% of deals
52% lost
48% won
Reached UpsideEach square is 1% of deals
79% lost
19% won
2% open
Deals that reached the Commit or Upside forecast category, 2022 to 2026; a deal can appear in both. Average win rate across all closed deals: 12%. Source: AeolusGTM CRM analysis.
Reps will just progress deals even though the deal doesn't need to be progressed.
Andy Smidmore, RevOps leader, on the RevOps Lab podcast, May 2026
"All it does is give false hope to leadership," Smidmore added. When a pipeline review rewards movement, reps supply it, and the CRM records that movement as history the next model treats as fact.
Check yoursPull every deal that reached Commit in the last four quarters and the share that closed lost. If nobody can produce that number within a day, the history an AI forecast needs isn't being kept.
Close dates
Finding 02 / 06
First forecasts almost never held, and close dates slipped further every year.
97%
of won deals missed the month they were first forecast.Two in three deals had their close date pushed at least once.
A close date is the single field every forecast leans on, and in this history it rarely survived contact with the buyer. Nearly every won deal closed in a different month from the one first forecast, and two-thirds of all deals had their date pushed at least once.
The drift got worse over time. Deals created in 2025 averaged 143% more pushes than deals created in 2020, even though newer deals have had less time to slip. A model trained on six years of this learns to treat a moving date as normal, and stops reading the close date as a signal.
Close-date pushes per deal more than doubled between 2020 and 2025
Average number of times a deal's close date moved into a later month, by the year the deal was created. Same-month moves are not counted, so these are floors. Source: AeolusGTM CRM analysis.
54% of won deals and 45% of open deals were dated on day 25 or later
Close dates by day of month, as a share of each group. The dashed line is the share an even spread would give, about 8% on days 29 to 31. Lost deals cluster less, with 19% on day 25 or later. Source: AeolusGTM CRM analysis.
Some month-end clustering in won deals reflects when buyers sign. Open deals carry dates a rep chose, and 45% of them sat on day 25 or later, which points to dates set to fit a reporting period rather than a buyer's timeline.
Stage probabilities told a different story, landing within six points of real win rates at every stage. The same CRM held trustworthy history and unreliable history side by side, and only testing shows which is which.
Check yoursFind the won deals that recorded a first-forecast date and count how many closed in that month. The answer tells you how much weight a forecast date deserves.
Pipeline age
Finding 03 / 06
68%
of open pipeline value sat in deals open 1.5x longer than the average win takes to close.Only 8% of deals that old ever closed won.
Most pipeline had outlived the sales cycle, and deals that old won half as often.
The average won deal closed in about four months. Two-thirds of open pipeline value had been open longer than six, and history says those deals rarely come back: past 1.5 times the average win cycle, the win rate fell by half and never recovered.
The pipeline total looks healthy because nothing ever leaves it. An AI forecast counts every one of those deals, weighted by a stage that hasn't moved in months.
Check yoursTake your average won sales cycle, multiply it by 1.5, and add up the open pipeline older than that. That share is the part of your forecast built on deals your own history says will not close.
Win rates fell by half once a deal outlived 1.5x the average win cycle
Win rate of closed deals by sales-cycle length, as a multiple of the average won cycle. Source: AeolusGTM CRM analysis.
Nine in ten open deals carried at least one risk flag, and most carried two or more
93% of open deals carry at least one flag.
63%Two or more flags
30%One flag
7%No flags
Open deals with each flag
Next step untouched in 90 days67%
No logged activity in 90 days51%
Close date pushed 3 or more times42%
48% of open pipeline value
Close date already passed25%
Flags overlap, so the shares do not add up to 100%.
Each square is 1% of open deals at export, May 2026. Value share uses the Amount field on open deals. Source: AeolusGTM CRM analysis.
Buyer qualification
Finding 04 / 06
The fields that predict wins sat mostly empty, and the quality score never checked them.
Deals with an Economic Buyer named won 2.6 times as often as deals without one. The gap held at every stage of the pipeline, so it reflects more than late-stage deals carrying more detail. Deals with two or more buyer fields filled won 2.7 times as often as deals with none.
Only one deal in seven named an Economic Buyer, and fewer than one in fifty named a Decision Maker or a Champion. The CRM's own data-quality score graded Type, Lead Source and Next Steps instead, and the deals it rated perfect won less often than the deals it rated lowest.
Naming the Economic Buyer raised the win rate at every stage of the pipeline
Win rate of closed deals with and without an Economic Buyer recorded, overall and by the furthest stage reached. Stage results use deals with stage timestamps. Source: AeolusGTM CRM analysis.
Stage and close date were filled on every deal; the buyer fields on almost none
Share of opportunities with each field populated. Source: AeolusGTM CRM analysis, field profile across 19 CRM objects.
Required fields always fill, though not always with the truth. "Too often, Salesforce users get around validation rules with placeholder text like 'unknown', 'TBD', or a made-up value," Lars van Bergen wrote on Salesforce Ben in August 2026.
A completeness score treats every field as equal, and the win rate shows they aren't. A readiness check worth running finds which fields predict outcomes in your own history, then measures how often your team fills them.
Check yoursCompare the win rate of closed deals with an Economic Buyer recorded against those without. If the gap is large and the field is mostly empty, your best signal is the one your reps skip.
Loss reasons
Finding 05 / 06
Four in five lost deals gave no usable reason, so a model can't learn why you lose.
81%
of lost deals gave no usable reason.Only 4% named a competitor.
Deal-scoring models learn as much from losses as from wins. Salesforce's Einstein Opportunity Scoring won't build a model on an org's own data until it has 200 closed-lost deals from the last 24 months, and Gong's deal model reads two years of closed deals.
Here, 42% of losses were marked "No Decision," 23% "Other" and 16% were left blank. A model reading that history can learn that deals die, and almost nothing about why.
Only one lost deal in five recorded a reason a team could act on
Share of lost deals by recorded loss reason. Source: AeolusGTM CRM analysis.
Each model reads three to 24 months of closed deals, losses included
- 1Einstein Opportunity Scoring24 months200 won + 200 lost, plus stage history
- 2Gong deal likelihood24 monthsweighs close-date changes
- 3Einstein Forecasting12 monthsneeds an update in every month
- 4HubSpot AI forecasting3 monthsclosed-won dates and amounts
Sources: Salesforce, Opportunity Scoring considerations ยท Salesforce, Forecasting considerations ยท Gong, deal likelihood ยท HubSpot, AI projections. Accessed October 21, 2026.
Activity, the other input deal models read, told the same story. Nearly nine in ten logged activities were automated marketing sends, and most of the rest never attached to a deal, so the record of what reps actually did is thin exactly where a model looks for it.
Check yoursPull last year's lost deals and count the ones with a specific reason. If most say No Decision, Other or nothing at all, your losses are teaching a model nothing.
The checklists
Finding 06 / 06
AI can now edit deal history, and 96% of readiness checklists never test it.
HubSpot's Smart Deal Progression reads call transcripts "alongside the full deal history" and suggests changes to stage, amount, next steps and close date, and any property can be set to apply them automatically. Salesforce's pipeline agent makes the edits itself once Update Fields Autonomously is switched on.
We audited 52 public AI readiness checklists from 44 publishers, including Salesforce, HubSpot and 22 implementation partners, against 17 tests. They check access, integrations, completeness and duplicates thoroughly, and almost none test the deal history these tools learn from and now rewrite.
Of 52 readiness checklists
check access and permissions
check how complete fields are
test deal history
Hygiene tests fill most of the grid. Two checklists test deal history.
AeolusGTM audit of 52 public AI readiness checklists, October 2026. Filled squares: Clientell AI (stage history) and Revenue.io (stage history, and CRM against phone records). Hollow marks raise the idea without a test and count as no. Coded dataset (CSV)
Two checklists test history. Clientell's Agentforce checklist sets a "Stage skip rate below 15%." Revenue.io checks stage history and is the only item that holds the CRM against an outside record, comparing logged calls with the phone system.
Eight more raise a history question without testing it. One asks whether close dates are "treated as forecast inputs." Another asks whether Commit drifts "with rep optimism and manager pressure." The tools that read live CRM data all read its current state, which shows none of the history a deal model trains on.
An AI edit can become the next model's evidence
How are things looking for the renewal?
Probably next quarter.
| Suggested | From → to |
|---|---|
| Close date | Mar 31 → Jun 30 |
| Stage | Demo → Proposal |
| Close date | Auto |
| Stage | Auto |
| Amount | Review |
Does the history keep who changed the field?
| Field | Value | Changed by |
|---|---|---|
| Close date | Jun 30 | Rep or AI? |
| Stage | Proposal | Rep or AI? |
| Signal | Records read |
|---|---|
| Stage changes | |
| Close date changes | |
| Amount changes |
| Deal | Close | Stage |
|---|---|---|
| Jun 30 | Proposal | |
conversation
Illustration; field values are examples. We found no published accuracy rate for AI-suggested deal updates from either vendor. Sources: HubSpot Spring 2026 Spotlight; HubSpot Knowledge Base, recommended next steps for deals; Salesforce, Agentforce for Sales Pipeline Management Implementation Guide.
CRM AI tools are only as useful as the data they are trained on.
The standard
What to test
Five tests that show whether your CRM is ready for AI.
If you sell AI readiness, add these. If you buy it, ask for them by name. Each one runs on history your CRM already keeps, and the last column shows how this dataset scored.
| Test | What it measures | How this dataset scored |
|---|---|---|
| Commit outcomes | Share of deals that reached Commit and then closed lost. | Fail52% lost |
| First-forecast accuracy | Share of won deals that closed in the month they were first forecast. | Fail3% held |
| Pipeline age | Share of open pipeline value older than 1.5x your average won cycle. | Fail68% of value |
| Buyer-field signal | Win rate with and without key buyer fields recorded, and how often they're filled. | Signal found2.6x lift, 15% filled |
| Loss-reason quality | Share of lost deals with a specific, actionable reason. | Fail19% usable |
One input passed: stage probabilities landed within six points of real win rates at every stage, which shows a CRM can hold trustworthy and unreliable history at the same time.
What to do
This quarter
Three moves before an AI forecast reads your CRM.
Keep AI off the forecast fields.
In HubSpot, leave the Automated box unchecked for close date, stage and amount. In Salesforce, leave Update Fields Autonomously inactive. Check a month of AI suggestions against what closed before you change either.
Measure what Commit is worth.
Pull the share of Commit deals from the last four quarters that closed lost. In this dataset it was 52%. If it takes more than a day to produce, fix that before a model reads the same records.
Find the fields that predict your wins.
Compare win rates with and without Economic Buyer, Budget and Decision Maker recorded. Require the ones that move win rate at the stage where reps learn them.
The AeolusGTM CRM Diagnostic
We run all five tests on your CRM and show you which records AI can trust.
It connects read-only to HubSpot and scores data quality, pipeline hygiene and the history tests in this report: Commit outcomes, first-forecast accuracy, pipeline age, buyer-field signal and loss-reason quality.
Then we walk you through the deals behind each result and show you which fields predict wins in your own history, which records an AI forecast should ignore, and what to fix first.
- You connect
- HubSpot, read-only. For any other CRM, we work from an export.
- You receive
- A scored report, a live walkthrough and a ranked list of what to change before you trust an AI forecast.
- If it's worth doing
- That list becomes the scope of the work. If not, the report is yours.
Questions
What this report answers
Questions this report answers.
What is the 2026 AI Readiness Audit?
The 2026 AI Readiness Audit is AeolusGTM’s October 2026 research report on whether CRM data is ready for AI forecasting and deal scoring. It tests six years of anonymized B2B CRM deal history (1,601 opportunities, 733,834 records and 1,021,915 pipeline snapshots) and audits 52 public AI readiness checklists from 44 publishers against 17 tests. The coded checklist dataset is public.
Do AI readiness checklists test CRM deal history?
Almost never. Of the 52 public AI readiness checklists AeolusGTM audited in October 2026, 96% never test deal history: how often close dates moved, which stages deals reached and how they closed. That history is what AI forecasting and deal-scoring tools learn from. 79% check access and permissions and 71% check how complete fields are. Only Clientell’s Agentforce checklist and Revenue.io test stage history.
How can I tell if my CRM is ready for AI?
Test the deal history your CRM already keeps. The report sets out five tests: the share of Commit deals that closed lost, the share of won deals that closed in their first-forecast month, the share of open pipeline value older than 1.5 times your average won cycle, win rates with and without buyer fields such as Economic Buyer, and the share of lost deals with a specific loss reason. For a broader self-assessment that also scores the people using the AI, see our AI readiness assessment for sales teams.
How much CRM history do AI forecasting tools learn from?
Between three and 24 months of closed deals, depending on the tool. According to vendor documentation read on October 21, 2026, Salesforce Einstein Opportunity Scoring reads 24 months and needs 200 won and 200 lost deals plus stage history, Gong’s deal likelihood model reads 24 months and weighs close-date changes, Einstein Forecasting reads 12 months, and HubSpot’s AI forecasting reads 3 months of closed-won dates and amounts.
How reliable are CRM forecast categories and close dates?
In the CRM data behind the report, 52% of deals that reached Commit closed lost, and 97% of won deals with a recorded first-forecast date closed in a different month from the one first forecast. Stage probabilities held up, landing within six points of actual win rates at every stage. These results come from an anonymized B2B CRM dataset, so test your own history before an AI forecast learns from it.
Which CRM fields best predict whether a deal will win?
In the report’s dataset, the Economic Buyer field was the one most tied to winning. Deals with an Economic Buyer named won 2.6 times as often as deals without one, and the gap held at every stage of the pipeline. Deals with two or more buyer fields filled won 2.7 times as often as deals with none. Only 15% of deals named an Economic Buyer, while stage and close date were filled on every deal.
Should AI update CRM close dates and stages automatically?
Not before you have checked its suggestions. HubSpot’s Smart Deal Progression suggests changes to stage, amount, next steps and close date, and any property can be set to apply them automatically. Salesforce’s pipeline agent edits fields itself once Update Fields Autonomously is on. The report recommends leaving automation off for close date, stage and amount until a month of AI suggestions has been checked against what closed. We found no published accuracy rate for AI-suggested deal updates from either vendor.
Methodology
How we built this report
- Checklist audit
- In October 2026 we searched the web for public AI readiness checklists, assessments and guides that tell CRM and revenue teams what to check before using AI. 52 items from 44 publishers qualified. Each was coded yes or no on 12 hygiene dimensions and five history tests, listed in the CSV. A test counts only when the item contains an explicit check; eight borderline items count as no. Seven gated items were coded on their public portion.
- Vendor documentation
- Read October 21, 2026, from Salesforce, HubSpot and Gong help centers, implementation guides and newsrooms.
- CRM data
- An anonymized B2B CRM dataset covering March 2020 to April 2026: 1,601 opportunities, 733,834 records and 1,021,915 pipeline snapshots. Win rates use closed opportunities. Stage results use opportunities with stage timestamps. Commit and Upside cover 2022 onward. Sales-cycle comparisons use the average won cycle. First-forecast accuracy uses won deals with a recorded first-forecast date. Pushes count close-date moves into a later month. Open-pipeline figures are as of May 2026.
- Our interest
- AeolusGTM sells CRM diagnostics and revenue-system work, and the CRM data comes from that work. Results are descriptive; we ran no significance tests. The checklist dataset is public so you can check every call we made. If we miscoded yours, message Sam on LinkedIn and we'll correct it.
- Cite this report
- AeolusGTM, The 2026 AI Readiness Audit, October 2026, https://aeolusgtm.com/ai-readiness/. The checklist dataset is free to reuse under CC BY 4.0 with attribution.