Most teams start with one person on Operations Workbench Starter, free for 30 days. Add LevelOneAI in Claude or ChatGPT, connect your email and add any exports you have, such as open POs or a schedule. Starter shows which changes put orders at risk, with the evidence, and tracks each one to a close. When the whole team needs it, with live ERP data, approvals and write-back, that's the Operations Workbench.

How do Starter, the Workbench and the Operations Engine fit together?

The LevelOneAI Operations Engine does the work: it treats each order, PO or job as one record (Work), compares what changed with what was promised (Context), works out what it affects and who needs to decide (Decision), and assigns, drafts and tracks the next step (Action), until it's resolved. Starter runs the engine on one person's email and the documents they add. A Workbench runs it across a whole team's email, ERP and CRM for one workflow, with live data, approvals and write-back.

Is there a free trial?

Yes. Operations Workbench Starter is free for 30 days, and you don't need a card. After that it's $199/month or $1,990/year (two months free), for 1 user and 1 mailbox. If you don't continue, your account becomes read-only and your data is deleted 30 days later.

What do I need to use Starter?

A paid Claude or ChatGPT plan (the free tiers don't include the access it needs) and a Gmail or Outlook mailbox. For work email, a business or enterprise plan from that provider is recommended. Documents are optional: exports from any ERP or spreadsheet work.

What's the difference between Starter and the Operations Workbench?

Starter works from one person's email and the exports they add, and you set it up yourself in minutes. It's built to show the value on your own work. The Operations Workbench connects live to your ERP and planning data, runs on its own, puts the whole team in one queue, and adds approvals, drafted follow-ups and write-back. Starter fees paid in the 6 months before you buy a Workbench are credited toward its setup.

Which documents should I add to Starter?

Whatever you have, or none. It works on email from the first run, and tells you which document would sharpen the results most. Typical ones are an open-PO export, a production schedule, open sales orders or an AR aging report. It keeps only the fields it needs from them, encrypted, never the files.

It says what it doesn't know. Each fact shows its source and when it arrived. When an email could match more than one order, it asks instead of guessing, and when a document is missing it says "impact not assessed" rather than "no impact." Your team makes the call.

How does billing work?

Starter is billed monthly or annually by card, with a receipt emailed for every payment and invoices on your billing page. Cancel anytime. A Workbench is a one-time setup fee, invoiced on the schedule in your quote, plus an annual platform fee that starts at go-live.

When does my team need a Workbench?

When the work depends on more than one person's inbox: the evidence sits in your ERP, CRM or planning data, several people need to see the same item, and someone has to make sure it closes. A Workbench gives the team one queue, one record per item, approvals and a team view.

What's built, and what do you configure for us?

The Operations Engine, the work queue, the work record, approvals, follow-up tracking and the team view are built. For each Workbench we configure the connections to your systems, the rules that decide what counts as an exception, who owns what, what closes an item, and whether approved changes are written back. Which systems it connects to, and how, is worked out during setup.

Operations Workbench Starter is free for 30 days, then $199/month or $1,990/year. An Operations Workbench is a one-time setup fee, from $20,000 per workflow, plus an annual platform fee that's included in your quote. There's no ongoing services fee: once it's live, your team runs it, and our help is optional. The platform fee covers hosting, AI processing, connector updates and improvements to the Operations Engine. It starts at go-live, and in the first year you can cancel 90 days after go-live if the agreed measures aren't met. The payback calculator compares the setup fee with what the delay costs you each year. The next Workbench reuses the connections you already have, so it costs less to set up. See Pricing for what's included.

How long does setup take?

Starter works the same day. A Workbench typically goes live in 6 to 12 weeks, depending on how many systems it connects and how complex the rules are. We agree the baseline and what success looks like before setup starts.

What happens to our ERP and CRM?

They stay where they are and keep the master data. A Workbench reads from them. Write-back is optional: whether approved updates go back, and to which systems, is decided during setup. Nothing changes in your systems without approval.

What data does LevelOneAI keep?

It's designed to keep only what it needs to run the work, such as references to the source, the facts it extracted (a date, a quantity, a price), status, owners and decisions. For a Workbench, exactly what's stored is agreed with you during setup. Starter never stores email: it keeps short labels, plus the fields it needs from documents you add, encrypted.

What will this ask of our IT team?

Starter needs no IT: you connect your own email in Claude or ChatGPT, and there's a one-page security summary to share with IT. A Workbench needs read-only access to the systems the workflow uses, granted once by whoever administers them, and typically a few hours of IT time during setup. Nothing is installed, and nothing changes in your systems without approval. See Growing Businesses if you don't have an IT department, or Enterprises for how we fit an IT review.

What happens when it isn't sure, or data is missing?

If an email could belong to more than one order or PO, it shows the candidates with their evidence and asks a person to confirm, instead of guessing. If a source is unavailable or a record can't be found, the item is marked incomplete, so your team knows the picture isn't complete. Every fact is labeled confirmed, conflicting, inferred or unknown.

Is this a product or a consulting engagement?

A product. The Operations Engine, the work queue, the work record and approvals are built. Setup configures them to your systems, rules and owners for one workflow. You pay a one-time setup fee and an annual platform fee. There's no ongoing services fee, so your team runs it without us.

Do you have a SOC 2 report?

No SOC 2 report yet. We'll walk your security team through the data flow and complete your security questionnaire on request.

What happens to our data if we stop?

You can export your work history and audit trail. With Starter, your data is deleted 30 days after you cancel or after a trial ends without a paid plan. Your ERP and CRM keep their own data, because they were the record all along.

That's decided with you during scoping, based on your security and IT requirements.

To read the messy information around a piece of work (the email, the attachment, the system field), match it to the right order or PO, compare it with what was promised, and recommend the next step. Every fact is labeled confirmed, conflicting, inferred or unknown, with its source. Your people make the call, and nothing changes in a system without approval.

Why can't we just add an AI agent?

You can, and it will do the steps. What an agent won't do on its own is keep one record of each order across email and your systems, notice a change nobody asked about, know which promise it breaks, and keep the issue open until it's resolved. That's what you'd have to build and maintain around it. LevelOneAI already does that, and works alongside the agents you use. See Why AI Agents Fail under Resources.

How is this different from process mining?

Process mining reads event logs from your ERP or CRM, so it can't see what happens in email, spreadsheets and phone calls, which is where most exceptions start. It also tells you where the process breaks; it doesn't work each exception to a close. A Workbench does.

Why not wait for our ERP or CRM vendor to add AI?

Use it. ERP AI is getting good at work inside the ERP, such as reading a supplier's email and proposing the PO change. A Workbench is for what happens next, across systems: which jobs and customer orders the change puts at risk, what the planner's sheet says, who has to approve, and whether it was actually resolved. If the problem starts and ends inside your ERP, its own AI may be all you need.

When is our ERP's own AI enough?

When the problem starts and ends inside the ERP, updating the record is the whole job, and everything the decision needs is already there. Add a Workbench when the change matters because of what it affects elsewhere (jobs, customer orders, invoices), when the numbers people trust live in a spreadsheet or portal, when resolving it takes several people and an approval, or when someone has to confirm it was actually resolved.

What if our workflow changes?

Rules, owners, thresholds and closing conditions are configuration, so they can be updated as your process changes, without a new project.

What if our team doesn't adopt it?

A Workbench brings the context to each piece of work instead of asking people to go and look for it, and it fits around the systems they already use. We also agree the measures up front, so you can see whether it's working.

All insights

General

The Intelligence Can Be Probabilistic. The Workflow Shouldn't Be.

By Shanshan "Sasha" Ma

LLMs are probabilistic — which is exactly why they belong in the information work, and not in the part that changes the state of the business. Why so many agents stall the moment they touch real orders, invoices, and schedules, and how to split the work so probabilistic AI improves human decisions while deterministic software executes them.

You've probably heard this one: LLMs are probabilistic, so the best way to use them is in probabilistic environments — creative work, exploration, open-ended reasoning, summarization.

It's a good rule. It's also where a lot of AI programs stop thinking.

Where the claim is right

When AI is generating information, probabilistic is a feature, not a risk. Ask for ten variations of a headline and you want ten different answers. Ask for a summary of a long thread and you want the one that captures the gist, not a hash of it. Ask for hypotheses, options, drafts, a first pass over messy text — the whole value is that the model explores instead of returning a lookup.

Nobody audits a brainstorm. Nobody needs the same email draft twice.

Business operations have a different requirement

But operations aren't a brainstorm. The closer AI gets to changing the state of the business, the less tolerance there is for probabilistic behaviour — and it isn't about model quality, it's about consequences.

Generating information leaves the world exactly as it was. Changing an order, approving an invoice, releasing a schedule, updating a record — those don't leave the world as it was. They move money, commitments, and customers.

The intelligence can be probabilistic. The workflow should be deterministic.

That sentence is the whole argument. Intelligence is how you figure out what should happen. The workflow is how it actually happens — and that part should behave like software: predictable, repeatable, auditable.

Why so many agents struggle in business environments

Because business processes accumulate consequences.

An error in a draft costs you the time to fix the draft. An error in an executed action costs you an order that has to be corrected, a payment that has to be clawed back, a customer who was told the wrong date, a schedule that has to be re-planned across three teams, a financial record that no longer reconciles with anything.

And it compounds. An agent that acts on a slightly wrong state, and then acts again, and again, doesn't produce one error — it produces a drift nobody notices until month end.

A 5% error rate in generating marketing copy might be acceptable. A 5% error rate in any of these is not:

Changing an order
Approving an invoice
Modifying a contract
Notifying a customer
Changing a production schedule
Updating a financial record

The 5% didn't change. The blast radius did. Most agent projects that stall in a business environment don't stall because the model was bad — they stall because the model was handed write access before anyone designed what “done” means.

Put AI in the right place

None of this means AI shouldn't be involved. It means the involvement should be designed. Three parts, in this order:

AI gathers and interprets the information. This is the messy, probabilistic work — and it's the work humans are worst at doing at scale: reading the thread, the ERP record, the supplier portal, the contract PDF, and three conflicting spreadsheets, then assembling what's actually true right now, including the exception that changed this morning. This is where probabilistic intelligence earns its place, because the input is ambiguous and the output is information, not an action.

A human makes the judgment. Commitments, money, risk, exceptions, anything customer-impacting. The person decides — but they decide from a state someone else assembled, in seconds, instead of rebuilding the picture from six tabs.

A deterministic workflow executes the outcome. Once the decision is made, the execution is software: the same steps every time, with the approval recorded, the systems updated, the notifications sent, and any failure surfaced as work rather than swallowed.

AI does the understanding. The human does the deciding and carries the accountability. The workflow makes sure the decision is actually executed.

What deterministic actually buys you

The same decision produces the same action. Not usually the same — the same.
Retries are safe. A step can fail and re-run without ordering the part twice or paying the invoice twice.
Every action is auditable. Who decided, on what information, at what time, with what result.
Failures are visible. When a system is down, the work item says so, and it lands with an owner instead of disappearing into a log.
The judgment is preserved. The human's reasoning isn't re-derived by a model on the next run.

That's what “the workflow should be deterministic” means in practice. Not that nothing goes wrong — that when something goes wrong, it's knowable, reversible, and owned.

The goal isn't probabilistic business decisions

It's the opposite, and it's worth being precise about: use probabilistic AI to make human decisions better, and deterministic software to make those decisions reliably executable.

That's a different design than “give the agent the tools and let it run.” It means deciding, step by step, which work is interpretive (AI), which is judgment (human), which is deterministic (software), and which shouldn't exist at all. Then building the workflow that holds them in that order, with the human checkpoints sitting exactly where the consequences are.

Get that order right and the model's unpredictability stops being a liability. It's the part of the system that's allowed to be uncertain, because nothing it does changes the state of the business on its own.

Where LevelOneAI fits

The pattern is usually the same: the agent works beautifully in a demo, and nobody will let it near a real order. That isn't a model problem. It's an operating-model problem.

A LevelOneAI Operations Workbench is built in that order: AI assembles the state from your email and systems, a person makes the call, and the follow-through runs as tracked work until it's resolved. Nothing changes in your systems without approval. See how the Operations Engine works, or try Workbench Starter free for 30 days.

The intelligence can be probabilistic. The workflow shouldn't be.

Talk to us