Deals
The deal, as the record actually has it.
Ask Maddox about a deal and it reads that deal: its stage, its qualification, the objections still open, the meetings and documents attached to it, and what is known versus what is missing. Ask about your pipeline and it reads the stored snapshot, per stage, including the stages with nothing in them.
How it works
From a question to a grounded answer
The same three steps run whether you asked about one deal or the whole board.
It works out which deal
A name is resolved to a record you are allowed to open. If two deals could match, Maddox asks which one rather than picking, because a wrong read is a correction and a wrong write is damage.
It reads the record
Core facts, playbook position, qualification, open objections, stage grades, meeting and document evidence, and an explicit list of what is known and what is still a gap.
It answers with sources
Figures are quoted from what it read. Where a deal has never been reviewed, it says there is no review, rather than presenting an absence as a finding.
In the product
An answer you can check against the deal.
Worked example
How do I close this deal?
A normal CRM shows an AE the fields somebody filled in. MADDOX reads the deal, the meetings on it and what the customer actually said, then answers the question the AE actually has.
The question
An account executive types: “How do I close the [company] deal?” Maddox resolves the name to a deal they are allowed to open — asking which one if two could match — and reads before it answers.
Where the deal stands
First the deal’s own context: stage, playbook position, qualification and MEDDPICC, stage grades, and the deal sheet’s known facts and outstanding gaps. From the latest grade it takes the grader’s one thing to do next. From the gates it takes the exit criteria and move-forward checklist for the current stage, which is often the most direct answer to “what is standing between this and a signature”. If the deal has a stored review, its risks, gaps and blockers are quoted as recorded; if nobody has reviewed it, the answer says so.
What the customer has said
The buying signals come back as the customer’s own sentences, each with its tier, the meeting it was heard in and the date. Beside them are the close opportunities: each moment the customer handed the rep a chance to ask, what the rep said next, and the verdict on whether the close was taken. Moments the customer could not yet have meant are left out. Where there is no recording, or one that has not been read yet, the answer tells those apart rather than reporting that there are no signals.
The pains on the deal come back with how deeply each one has been explored and any impact the customer put on it. The five closing steps themselves live on the deal’s Path to Close panel; what Ask Maddox reads is the evidence underneath them — the pains, the facts, the gaps and the open objections.
What is in the way
Every objection the deal has raised, from the BDR’s first call to the AE’s latest meeting: who raised it, at which step, the quote, and whether it was beaten. An objection with no captured sentence is marked as such rather than read as silence, and any of them can be opened on the call itself, at the exchange where it happened.
Then the promises: what we owe the customer and what they owe us, with anything overdue. And if the deal has no next step, or an overdue one, that is named — both are among the reasons a deal is flagged as needing attention.
What the answer looks like
The reply is built from those reads: the gate item still unmet, the objection still open in the customer’s words, the close moment that went past, the commitment that is overdue, the next step that is missing. Each point links back to the deal, meeting or call it came from. It does not invent a win probability or re-score the deal on the spot; any figure it quotes is the one already stored, and where the rep’s own call, the evidence read and the model score differ, the three are reported separately rather than averaged.
Turning it into action, with your approval
Most deal conversations end with something to do, and the follow-up questions are where Maddox stops reading and starts proposing.
It will not move the stage, edit the deal or send anything on its own. It can read the whole deal; the only changes it can make from here are the task and the meeting, and only after you click approve.
- “Draft me a follow-up task to send [contact] the revised proposal on Thursday.” Maddox puts a task card up for approval, linked to the deal and assigned to you. Nothing is saved until you approve it.
- “Find a time for the next meeting with [contact] next week.” It checks the calendar and proposes a slot on a card; the meeting is booked only when you approve, and the slot is checked again at that moment. Without Microsoft 365 connected it can check only the calendar inside MADDOX, and it says so.
- “Which objections on this deal are still open?”
- “Show me the close moments on this deal and what I said next.”
- “What did the last review say?”
The mechanism
Four tools, and the work each one refuses to do
Deals are the area where a plausible answer is most expensive, because the numbers are the ones people commit to out loud. Each tool here is shaped around a specific way a fluent model gets a deal question wrong.
Ranking is a closed list of seven orderings, run in the database
Superlative questions — biggest, closing soonest, sitting longest — have exactly one correct answer, and it is not one a language model should work out. Handed a list of deals it will compare amounts and get it wrong on a tie, a null, or the eleventh row. So the ordering happens in Postgres, across your whole book, before a single row reaches the model. What comes back is already ordered, and the model can only narrate it.
The available orderings are a fixed enumeration rather than a free-text sort: largest and smallest by amount, soonest and latest close, oldest, newest, and least recently updated. The model cannot name a column, so it cannot reach one it should not see or write a query by suggestion.
Nulls are placed explicitly on every one of them. A deal with no amount is not a zero-value deal and a deal with no close date is not closing today — left to the database default, “closing soonest” would put every undated deal first. Rows missing the ranked value are excluded and counted, so the answer can tell you that three of your fourteen deals have no close date rather than quietly ranking them.
The attention list is your own book, and absent is not empty
What-needs-attention is deliberately scoped to deals you own, because it is a to-do list rather than a report — a manager asking about a rep’s book is a different question with a different permission behind it, and answering it here by accepting someone else’s name would hide a scope check where nobody would look for it. The ranking and the reasons behind each position come from the same builder the product itself uses, and both are passed through unchanged. When the list is empty it arrives with a stated reason, because “you have no deals needing attention” and “you have no open deals” read identically unless the assistant is told which it is.
A stored review is quoted, never merged
Asked what a review said, Maddox reads the stored review rather than the deal record, and it is told to quote it rather than re-summarise it. Where several reviews exist it will not blend them into one tidier answer — earlier snapshots are listed by identifier for you to ask for specifically. If nobody has ever reviewed the deal, the tool returns that as a state of its own, which is why the page says there is no review instead of presenting an absence as a finding.
Beyond the four: the drill-down, the evidence and the signals
Four tools answer the everyday deal questions, and three more areas sit beside them for the questions that go deeper. A drill-down reads the deal the way its own page reads it, block by block. The evidence area reads the stored facts, the open and resolved objections, the evidence behind each stage grade and the documents filed against the deal. And the buying-signal reader returns what the deal has been signalling — stored rows, each with where it came from — rather than a model’s impression of momentum.
Each of those is its own area for the same reason campaign changes is: a question about one objection does not need the whole pipeline attached to it, and an assistant handed fewer tools chooses between them better.
The whole team’s pipeline is a different question with a different gate
Total open value and the per-stage split are org-wide by construction, so they sit behind the management permission rather than the ordinary deal one, and they come from a snapshot compiled overnight and quoted with the date it was taken. The split covers every stage including the empty ones, which is the part a live count tends to drop. Asking that question is not the same as asking about your own deals, and the tools are kept apart so an owner asking about the whole team is never quietly answered about their own book.
Who it is for
For the person who owns the number
Reps use it to find the deals that need them today. Managers use it to ask the pipeline a question without building a report first.
- Account executives, on their own deals, before a call
- Sales managers, on the pipeline as recorded rather than as recalculated
- Anyone who has ever opened six tabs to answer one question
Related
Questions
The things people actually ask.
Which deals can I ask about?
The ones your permissions already allow you to open. The what-needs-attention view is deliberately scoped to deals you own, because it is a to-do list rather than a report.
Can it tell me which deal is biggest, or closing soonest?
Yes, and that ranking is deterministic rather than a judgement call. You can order your deals by amount, close date, age, or how long it has been since anything happened on them.
Does it re-score or re-forecast the pipeline when I ask?
No. It reads the stored snapshot and the stored forecast, which is what people actually committed to. Recomputing a number on the fly would produce a figure that disagrees with the one on the dashboard, and you would have no way to tell which was right.
Can Maddox tell me how to close a specific deal?
It can show you what stands in the way, from the record. It reads the exit gate, open objections, outstanding gaps, buying signals and close moments, each with its source, and quotes the grader’s recorded next action rather than inventing a strategy. It changes nothing on its own: a follow-up task or a meeting needs your approval.
What if the deal has no next step recorded?
That is one of the reasons a deal gets flagged as needing attention, alongside an overdue next step, an open objection, an unmet exit criterion and an overdue promise to the client.
Ask it about the deal you are worried about.
Bring the one you would raise in a pipeline review anyway, and see whether the record agrees with you.