Ungoverned AI
no limitsHand it a risky task and it just does it.
- Acts on unverified data and missing documentation
- Blows past materiality limits and usage caps
- No gate between "uncertain" and "posted"
- The failure is silent — it looks like success
you stay in control
You set the limits. Aleq earns the rest.
Aleq's autonomy is a confidence score that climbs as it proves itself on your books — inside limits you set. Below the limit it drafts and waits; above it, it posts and shows the receipt.
every action logged · reversible by a standard entry
a_19d4 May reconciliation · JE-12491 reversible
a_19c8 dunning · Haven Foods logged
a_1a02 AWS bill B-3041 · auto-posted reversible
a_1a07 Kestrel wire · awaiting CEO held
logged · hand your auditor a full history
the one rule
Skill times permission.
Everything Aleq does clears two bars: can it do the work, and has it earned the right to — with you, on your books. Not one or the other. Both, multiplied.
Skillhow good it is at the task
Permissionhow much it's proven, with you
Actionwhat actually happens
multiplication, not addition — anything times zero is zero. all the skill in the world, without earned trust, does nothing.
It can't be bullied.
Overrule Aleq and turn out right — your word weighs more next time. Turn out wrong — it weighs less. Authority here tracks accuracy, not seniority. Not even the boss can shout the books into agreeing.
Questions count as wins.
When Aleq checks with you instead of guessing, that counts as getting it right. It's never punished for asking — so it never bluffs to look confident.
Trust isn't tenure.
It doesn't earn autonomy by being installed a long time. It earns it by being proven right — task by task, on your books, with your sign-offs as the evidence.
and some things don't bend: your preferences it learns in days · your controls it questions rarely · the rules that protect your books, never
it's not all-or-nothing
Three states. You set the limits that move between them.
The mode you see per task isn't a switch you flip — it's a read-out of how much Aleq has proven itself, gated by the limits and thresholds you've set.
manual
Below your threshold. Aleq advises.
New, unproven, or outside your limits. It looks things up and drafts the work. You make every call and post it yourself.
assist default
Proven, but not yet trusted alone.
It prepares the entry and shows its work. One click to approve — or reject — before anything posts.
auto
Above your threshold. Aleq acts.
It does the work and posts it, logged. Reverse it any time with a standard reversing entry.
doing this by hand
Keeping all this under control by hand is a second full-time job.
Today most of it is held together by hand. The rule that one person can't both approve a bill and pay it sits in a spreadsheet someone forgets to update. Approvals get chased over email. The week before the auditor shows up, someone builds the evidence by hand. And when they ask "why did this post?" there's no record — just memory. Aleq takes that work off your plate. Not by asking you to trust it, but by proving every move.
- Aleq blocks the same person from both approving and posting a journal entry — you don't track it in a sheet
- Approvals are built in — no email threads to dig back up
- The audit evidence builds itself all month — not the night before
| Who can do what | tracked by hand in a sheet | stale |
| Approvals | chased over email | no trail |
| Audit evidence | built the week before | 38 hrs |
| "Why did this post?" | rebuilt from memory | no replay |
one controller · most of a week, every close
how Aleq learns
It watches, earns confidence, acts inside your limits, and leaves a log.
Aleq learns from what your team already does. What happens next teaches it. It earns your trust with proof — you don't hand it over up front.
1 · learns247 bank lines
the statement, the matches, the ledger, and how you've coded these before
2 · confidence climbsCode to Stripe Clearing
you coded it this way 412 times and never reversed it
3 · posts the routine245 posted, 2 held
routine matches go through; two big ones wait for you
4 · loggedreversible any time
a logged entry you can reverse; what happens next updates the confidence score
This isn't a chatbot answer — it's a confidence score, the proof behind it, your limit, a logged entry, and a standard way to reverse it.
one task at a time
Aleq earns each task on its own — never all at once.
When confidence on a task crosses your threshold, that task — and only that one — runs alone. Bank matching can run itself while every payment still waits for your OK. Each task has its own limit, and you set it.
AutoMatching the bank
runs on its own after 200+ clean matches · 18,420 runs · 98.7% right · routine cash only
AssistSending a payment
always waits for you · drafts the wire, shows the proof, holds for your OK — never runs alone
ManualClosing the month
gets the close ready · a person signs the month closed
| Task | Runs alone after | Confidence | Mode now |
|---|---|---|---|
| Matching the bank | 200 clean matches | 1.00 | Auto |
| Coding bills | 150 times · within 4% | 0.99 | Auto |
| Sending reminders | small ones, below your escalation line | 0.96 | Auto |
| Sending a payment | never alone · always waits for you | 0.92 | Assist |
| Closing the month | person only | — | Manual |
if confidence on a task drops, Aleq stops and asks you again
what it looks like when it holds
Held for your approval — not guessed at.
A reliable rule still stops when the amount, account, or risk is outside what you've approved. Then it asks a specific question instead of guessing.
Needs approval
$188,440 Kestrel wire
The early-pay discount is worth taking, but this is over your $100k auto-payment limit.
match confidence 92%your limit $100kmode Assist
held for sign-off · no money moved
logged · you can replay it
Every action is logged. Hand your auditor a replay.
Everything Aleq does gets a logged event you can reverse. It's not a screenshot of a dashboard. It's a real record your auditor can check on their own.
- Every action shows why it happened, what set it off, the rule it used, and what it changed
- Routine work goes through quietly; big items come to you first
- Closed months are locked
| Why | 247 lines tie to the bank to the penny |
| What set it off | difference = $0.00 · within the rules |
| Rule | P-RECON-DIFF-ZERO |
| What changed | posts JE-12491 · closes May |
reversible any time · logged
reversible · try it first
Reverse anything. Test it first.
Every entry Aleq posts can be reversed with a standard reversing entry, any time — the same mechanism your team already uses, not a special countdown window. You can also test any action on your real books first — it shows what would happen and changes nothing.
- Every entry can be reversed with an offsetting entry, logged the same way
- Test mode: see the entry and what it would do, with nothing saved
| approved | wire | $188,440 |
| paid early | discount captured | −$3,768 |
reversible with a standard entry, any time
TAMi · you can read every rule
It shows you every rule it follows — and why.
TAMi — The Aleq Mind — is the part of Aleq that tracks confidence per task from your team's own approvals and corrections. Above your threshold, it runs alone; below it, it drafts and asks. Every correction you make updates the score for next time.
- Every rule shows its confidence and what happens if it's wrong
- When you approve or reject, that teaches it for next time
| Rule | Confidence | Status |
|---|---|---|
| Stripe transfer ⇒ Stripe Clearing | 1.00 | runs alone |
| AWS bill in last 3 days, within 4% | 0.99 | runs alone |
| Kestrel early-pay discount always on | 0.92 | learning |
| Ironwood pays only after a reminder | 0.42 | asks first |
the evidence · eval runs
When the stakes go up, Aleq stops guessing. We have the run to prove it.
We ran 57 high-risk accounting tasks that leading AI models got wrong on their own. Then we ran the same 57 with Aleq in the loop — same models, the only change is what Aleq lets them do when the answer isn't certain.
Aleq
inside your limitsThe same task routes to the safe move.
- Asks before acting on incomplete evidence
- Proposes the entry and waits inside your limits
- Escalates the judgment call instead of forcing one
- Refuses the unsafe path rather than fake a result
84%
came out right with Aleq in the loop — the same models, ungoverned, got all 57 wrong48 of 57 tasks
75%
Aleq stopped and asked, or queued the entry for your approval, instead of posting something wrong43 of 57 tasks
0
payments left without a person signing off — the controller keeps the keysacross every run
Those 57 failed tasks break into 91 issues across four risk areas — one task often trips more than one. The bar is how many Aleq caught — stopped, asked, or queued for approval — instead of letting through.
Financial reporting, close & record integrity46 failures · caught below76%
Audit, compliance & internal controls24 failures · caught below81%
Professional escalation & communication10 failures · caught below89%
Revenue, billing, contract & covenant risk11 failures · caught below67%
From Aleq's internal eval runs — leading AI models (Claude, Codex, Gemini, Qwen) on the same tasks, with and without Aleq. Request the full methodology →
closed months lock
A closed month is locked — no quiet edits after the fact.
When a month closes, Aleq locks the period: postings, reversals, and edits to closed entries are blocked at the ledger level. Reopening a closed period is its own explicit, logged action — not something that happens by accident.
- Closed and locked periods block new postings and edits
- Reopening a period is a separate, logged step
- Export a full replay of the period's entries
| Entries locked | 8,412 | locked |
| Closed by | Kurt | 14:02 |
| Status | closed | no edits |
reopening is a separate, logged action
tested before release
It passes the test before it touches your books.
Before any new version of Aleq goes live, we test it on real close, collections, and reconciliation cases — and run a set of safety checks. It has to score at least 0.90 out of 1.0. If it doesn't pass, it doesn't ship.
- Each job is tested and graded · re-tested on every change
- Safety checks have to pass before release
- The accounting can't quietly go wrong — a failed test blocks the release
| month-end close | 0.96 | pass |
| collections | 0.94 | pass |
| reconciliation | 0.97 | pass |
| safety checks | — | enforced |
| must score 0.90+ | — | ships |
a setting vs. a standing
Anyone can switch AI on. Aleq has to earn it.
The same models power everything now. What differs is the permission model — autonomy someone configures on day one, or autonomy that's earned on your books and can be lost.
Autonomy as a standing
earned · provablePermission that's accumulated, per task, from outcomes that checked out.
- Earned one task at a time — never all at once
- One miss costs more trust than ten wins earn, by design
- Stops at your limits and asks a clear question
- Every action logged and reversible; closed months locked
Autonomy as a setting
configured · trustedA threshold someone typed in on day one, and a lot of hoping.
- The same permission everywhere — first-time vendor or hundredth
- Confidence that never goes down, because nothing measures it
- Nothing between "not sure" and "posted"
- When it's wrong, "the setting was on" is the whole record
where this goes
Hired cautious. Grows into the one who's ahead of you.
day oneThe new hire
Asks about everything. Drafts, shows its work, posts nothing alone — exactly how you'd want a new controller to start.
months inThe trusted hand
Runs what it's proven — reconciliation alone while payments still ask — and every question it asks makes the next one rarer.
the steady stateAhead of you
The board pack is drafted the morning after close. The reminder went out at day one past due. You stop assigning the work and start reviewing it.
FAQ
What people ask before trusting it.
Not blindly — trust should be earned per task, which is how Aleq is built. Autonomy is not a switch you flip; it is a confidence score each task earns from your team's own approvals and corrections. Below the threshold you set, Aleq drafts the work and waits; above it, it posts and shows the receipt. Every action is logged with the rule that triggered it, and every entry can be reversed with a standard reversing entry. You connect read-only, start everything in Assist mode, and let individual tasks move to Auto only after they have proven themselves on your books.
You decide how much to trust it.
Connect read-only. Start everything in Assist. Let a task move to Auto only when it's earned it.