AI business copilot
AI agent for small business: say the goal, approve the plan
Type what you want and Esplanda returns a plan of costed actions, checked against your limits, with one button to approve it.
The copilot and the approval queue are part of the free Esplanda POS. No setup fee and no contract. See Esplanda pricing.
Get me to $50,000 a week at 23% margin
Illustrative example
In short
Esplanda's AI business copilot turns a goal, such as a weekly revenue target with a margin floor, into a costed plan you approve with one click.
Underneath the button: limits you wrote once in plain language, a daily spending ceiling nothing can raise, a fresh check of the facts before every action, and a kill switch two taps from anywhere. One click is only a good idea when the thing behind it cannot surprise you.
- Goal to plan
- One-click approval
- Limits you write
- Facts re-checked
- Kill switch
The last mile
The problem is not missing information
An owner running a single location can already see almost everything. What is missing is the work between seeing it and doing something about it.
Everything visible, in seven places
Sales, stock, reviews, ad spend, the supplier portal, the delivery app, the email tool, none of which agree on what a week is.
The last mile takes a morning
Somebody has to read it all, pick the two things that matter this week, decide, and go do it in whichever system owns the button. It happens after closing, and it is the first thing skipped in a busy month.
A decision, already prepared
A dashboard adds to that job. A decision that arrives with its dollar figure, cost, risk and downside reduces it, so your part is the judgment rather than the assembly.
How it works
From a sentence to an approval, in seven steps
A decision made on Monday's forecast is not automatically valid on Wednesday. That is why the last step re-checks itself.
State the goal in plain language
Money and time: a weekly revenue target with a margin floor under it. It becomes a real goal with a target, a period and a floor, and it stays on the home screen until you change it.
See the gap first
The answer opens with the number: how far behind you are, how many days are left, the daily rate that closes it and the daily rate you are running.
Read the plan as a table
Four to eight actions, each with expected dollars, a confidence figure, a cost to run and a risk level. Every figure traces to a source you can open: a forecast, a measure, a stored result.
Read what doing nothing costs
Both sides are required by the system: an item that cannot state its downside cannot be created.
See the exceptions called out
A promotion that dips under your margin floor, a send above your recipient ceiling: surfaced before you approve, with the rule named, not buried three rows down.
Approve once
One button for the whole plan. Everything moves to approved and queued together, scheduled items show their send times, and there is no second confirmation dialog.
Watch it run, and re-check itself
Before each item runs, the facts it was approved on are verified again. If the forecast has moved or the gap has closed, it stops, comes back marked re-evaluated and shows what moved, rather than running quietly or canceling itself.
Guardrails
The guardrails are why one click is acceptable
The limits are written by you, enforced by the system rather than by the AI, and cannot be widened by the thing they constrain.
| Guardrail | What it does | Why it holds |
|---|---|---|
| Rules in plain language | Never discount more than 10%. Purchase orders over $2,500 come to me. No texts before 9am | Checked on every action, every time, including when nobody is watching |
| A hard daily spending ceiling | A dollar cap on what automation may spend in a day | It sits outside the rules: no rule change, no automation and no AI request can raise it |
| Rules only a person can change | Writing or amending a limit is a person-only operation | The permission simply does not exist for an automated actor, so it cannot be granted by mistake |
| Freshness checks | An action whose facts have changed since it was proposed comes back to you | Returns for re-approval showing the difference, rather than running or silently dying |
| Constraints instead of rejections | A send that would exceed a contact limit is narrowed rather than blocked, and the reduction is shown | A rule that only ever says no gets switched off. One that adjusts gets kept |
| High-risk items held from bulk approval | Approving nine at once approves eight and holds the risky one for individual review | Bulk approval exists so routine work moves, not so an unusual item slips through with it |
| A permanent record | Who or what changed anything, when, why and under which version of which rule | Written as the change happens, and nobody can edit it afterwards, Esplanda included |
| The kill switch | Two taps from any screen. Everything queued is canceled and a banner appears everywhere | Works with the AI layer completely unavailable, because it deliberately does not depend on it |
A plan is governed by its most restrictive item
If five actions could run unattended and the sixth needs you, the whole plan waits. Automation does not get to run four fifths of a plan and leave the awkward part in your inbox. The full set of limits, the stop button and what happens to your data if you leave: why Esplanda is safe to trust.
How much is done for you
Three levels, set per kind of work
The copilot, the approval queue, the briefing and the guardrails stay the same. What changes is how much arrives already prepared, and how much Esplanda may do inside your limits.
Advise
Items arrive as recommendations with their numbers and their downside, and you carry out the work.
Prepare
Items arrive ready to execute, and one click both approves and performs them.
Act inside limits
Routine work proceeds and appears in the record; anything above your thresholds still waits for you.
Running unattended is earned, not switched on. It is granted per kind of action, once that kind of action has a measured track record: enough executions, a high acceptance rate, a low reversal rate, and results that landed near what was predicted. Asking for the highest level leaves that requirement in place.
Waiting never turns into doing. Anything that needs you and does not get you expires. It never proceeds after a timeout.
Every morning
Six blocks, a briefing already written
The home screen is six blocks and nothing else: no chart, no feed, no notification count. Opening Esplanda is meant to take ninety seconds.
- GoalWhat you asked for, in money and time, and when you set it.
- TrajectoryWhere the week is projected to land, the daily rate that closes the gap, and the daily rate you are running.
- What Esplanda is doingThe actions taken since you last looked, each with its measured result, and which of them needed nobody.
- What changedThe one thing that moved in the business, and the most likely reason, checked against something outside the business where one exists.
- What needs youThe items awaiting a decision and their combined dollar impact. Usually two or three; some weeks, none.
- Expected financial impactWhat approving everything queued is worth, and what approving nothing costs.
The block most software leaves out
If approving everything queued is worth very little this week, block 06 says so plainly, and an owner who reads that and closes the app has been told the truth.
Quiet week, quiet briefing
The briefing is generated before your delivery time. A quiet week gets a headline saying so, three facts and one line of what ran, rather than paragraphs written to justify the notification.
The numbers are the briefing
If the AI layer is down when the briefing is built, every number still arrives from a fixed template, marked as reduced, and the email still sends. An answer whose figure cannot be traced to a source is rebuilt, or sent as the plain templated version with a note. The deeper POS reporting is one drill-down away.
Ask why, get sources
Why was Tuesday bad? Am I going to hit my goal? The same question asked three ways returns identical figures, each a link to the rows behind it. Asking is covered in depth by the AI business assistant and AI retail analytics.
In your own language
Written in the language you set rather than translated out of English. Customer messages, reviews and supplier emails arrive readable, with the original one tap away. See multi-language business software.
Outside messages are read, never obeyed
A supplier email with text that tries to give orders is summarized, and the attempt is disclosed to you. No inbound message can approve anything.
When it learns something about your business, such as texts converting better late afternoon, it offers to remember it. It becomes a stored preference only when you confirm, and next week's plan respects it.
Where the plan comes from
The copilot assembles what Esplanda already found
Each finding arrives as a proposed action with a dollar figure. The copilot's job is ranking, costing, checking them against your limits and presenting the shortlist.
Forecasts, margin, marketing and reviews
The reorder that prevents a stockout, from demand forecasting software, the repricing on items sitting too long, the customer segment worth contacting this week from marketing, and the reviews and listings that need a reply.
Scenarios and your own rules
A modelled change from what-if scenario planning software becomes real changes with one approval, and a scenario over a week old is re-run first. Conditions you set in business automation software feed the same queue.
Measured afterwards
Every executed action is stored with what it was predicted to do and what it did. That record is the outcome ledger, and the misses stay in.
How to judge any system that claims to act for you: the levels of autonomy in business software. The whole loop, end to end: how Esplanda works.
Questions
Copilot questions owners ask
What is a business copilot?
A business copilot is a single place to ask about the business in plain language, get an answer built from the business's own numbers, and turn a goal into work that is ready to approve. Esplanda's copilot takes a sentence such as a weekly revenue target with a margin floor and returns a plan of four to eight costed actions with one approve button, rather than a dashboard for you to interpret.
Is it safe to approve a plan with one click?
It is safe because of what sits underneath the button. Limits you write once in plain language are checked on every action, including automated ones. A hard daily spending ceiling cannot be raised by a rule change, by automation or by the AI. Any action whose facts have moved since it was proposed stops and comes back for re-approval. Every change is recorded permanently. And a kill switch two taps from anywhere stops everything.
What is the kill switch and how fast is it?
Two taps from any screen: pause, then confirm. Queued automated actions are canceled, anything mid-flight stops at its next safe point, and a banner appears on every screen until it is lifted. The copilot still answers questions while paused, and acts on nothing. The switch works even when the AI layer is unavailable, because it deliberately does not depend on it.
Can Esplanda raise its own spending limit?
No. The daily ceiling sits outside the rules the system works with, so no automation, no rule change and no AI request can lift it. Separately, writing or amending the rules is something only a person can do, and every change is recorded with who made it and when.
What happens if the numbers change between proposal and execution?
The action stops. Every proposed action records the facts it was justified by, and before it runs those facts are checked again. If the forecast has been superseded or the gap has closed, the item returns to the queue marked as re-evaluated and shows what changed since you approved it. Esplanda never sends a hot-weather promotion on a cool day because the decision was made yesterday.
Can Esplanda act without asking me at all?
Only where you have allowed it, only for that specific kind of action, and only once that kind of action has a measured track record behind it. Running unattended is earned from evidence rather than switched on in settings, and it is granted per action type. Anything above the thresholds you set waits for you, and expires rather than proceeding.
What stops a supplier email from telling Esplanda what to do?
Anything that arrives from outside, such as a supplier email, a customer message or a review, is treated as content to be read, never as instructions to be followed. If a message contains text attempting to give the system instructions, it is summarized and the attempt is disclosed to you. No inbound message can approve an action, and no action approves itself.
What does the AI business copilot cost?
The copilot and the approval queue are part of the free Esplanda POS rather than an add-on. There is no setup fee, no contract and no per-user fee. Paid tools, such as marketing or the AI phone, are separate lines you add only when you want them. Every line is on Esplanda pricing.
Ninety seconds, six blocks, one button
Start free today, or book a demo and we'll turn one of your own goals into a plan, with your limits underneath it.
Esplanda POS and your website are free to start, with no card needed. Add paid tools only when you want them.
