Your ERP can only reason about what somebody keyed into it
Which is the whole problem with AI on a shop floor. The model is rarely the limitation. The data is. This page is what the intelligence layer does today, what it will not do, and the rules it operates under, because on this subject the rules matter more than the features.
Six rules this layer operates under
Every vendor in manufacturing software has an AI page. Almost none of them will tell you what happens when the model is wrong. These are the rules, and they are engineered rather than promised, which means you can test them.
The model never does the arithmetic
This is the design rule the whole layer rests on. The language model may plan the query. The runtime computes the answer deterministically against SyteLine, and the model only narrates the rows that came back. A number on the screen was calculated, not generated.
It abstains rather than guesses
Extracted values are validated against real records before they count. A page that does not resolve to a record stays unmatched rather than being filed somewhere plausible, and the attempted value is shown so you can see what it tried. Nothing is committed silently.
Every answer can be taken apart
Row level citations, plus a control that re-runs the underlying query and shows you the raw rows. If you do not believe the answer, you do not have to argue with it. You can open it.
A human reviews anything customer facing
Not a preference, a rule. Speed amplifies risk, and an AI output that reaches a customer without a person having read it is a category of mistake nobody wants to explain.
It waits for your history before it makes claims
Control limits need twenty to twenty five operating days of your own data. Until then the system shows targets and says so, rather than inventing a baseline. Demonstration data is never presented as yours.
Nothing ships before it clears an accuracy gate
Every customer facing AI claim has to pass an internal accuracy bar before it is announced. That is why this page is shorter than it could be, and why the things on it are dated.
The runtime computes against SyteLine deterministically and the model narrates the rows that came back. A number on your screen was calculated, not generated. That single design decision is why this layer can be pointed at a controller and a job cost rather than at a marketing page.
Four things you can use now
These four are available now. Everything further down this page is in development or planned, with a year where there is one.
Search or Ask AI across the documents on a record
Available nowDoc-TrakDoc-Trak FLEX follows you through SyteLine and shows the attachments for whatever record you are on. The Search or Ask AI bar answers a plain-language question across the attachments on screen, and the answer names the files it came from, so you can open the source rather than take the summary on faith.
It cites the files it used. An answer you cannot trace is not an answer.
AI summaries of any attachment
Available nowDoc-TrakEvery document attached to a record can carry a generated summary, so a twelve-page supplier agreement or a multi-page certificate can be understood without opening it. The summary sits on the record beside the file rather than in a separate tool.
Read the summary, open the source. Both are on the record.
Paperwork that files itself by rule
Available nowDoc-TrakSaving rules build the path, the name and the document type from the record's own fields. Mass Scanning matches a stack by barcode to receipts, lots, serials and vouchers, and Receive with Docs files the vendor's packing slip onto the PO as part of receiving. Nothing is filed somewhere plausible; a page that does not match stays unmatched.
The rule is the instruction. Nobody writes a prompt.
Triggers and alerts from the machine
Available nowMachine-TrakOut-of-cycle time opens a delay event by itself, triggers and alerts fire as a condition happens, and maintenance can see how often that fault has hit this machine and others. Top alerts and events by count and by duration, from records rather than memory.
The machine tells you when it stops, not the end-of-shift guess.
And one that is about us rather than you
Lake support runs on this. Draft answers are generated from a knowledge base built out of real cases, and an engineer reviews every one before it goes out. We would rather run on AI before selling it to you, and if it were not working we would be the first to know.
The question every defense and regulated shop asks first, answered plainly
Including the part most vendors leave out, which is what is not ready yet.
Your data is never training data
No SyteLine data, no customer data and no procedure content is used to train any foundation model. That is a contractual position, not a preference.
Read only, and queried in place
The layer reaches SyteLine through a read only role on an explicit allowlist, with a statement timeout and a row cap. It never writes to SyteLine, and there is no copy of your database sitting anywhere else.
Opt in, never automatic
Document intelligence sends nothing until somebody presses the button. Only the page image and the rule driven instruction leave. Your file system and your SyteLine data stay where they are.
Whose engine, honestly
Today Lake provides and maintains the model access. Customer-owned keys and a model running inside your own environment, which is what a controlled unclassified information environment needs, are a later phase and are not available yet. If somebody tells you otherwise, ask them to show you.
Regulated and defense environments
The machine intelligence collector runs on premise and supports regulated environments, and most of that install base is on premise today. The current AI work deliberately excludes the most restricted government tenants rather than pretending to cover them.
Planned, and not yet certified
Tenant isolation inside AI context, audit logging of AI output, role based access for AI features and a readiness assessment are all planned work. None of them is an attestation today and none of them is claimed as one.
What is coming, and how confident we are
AI across the platform flags risk
Planned 2027Jobs at risk of going late and cost drifting from standard, flagged across the platform before they become a late job or a margin miss. A person decides.
Ask the floor in plain language
Planned 2027A plain-language query over your SyteLine data, starting with standard against actual variance by job and operation. It runs on the AI engine you sanction, under your governance.
Every number carrying where it came from
In developmentData trust tiers surfaced inside the answer: system-captured signals you can act on, values a person entered that you verify before a major decision, and derived figures whose fidelity is bounded by the weakest input underneath them.
Cost variance, decomposed
In developmentNot that margin eroded, which a cost report can already tell you, but where it went: setup overrun from the labor record, absorbed downtime from the machine record, rework that was never tied back to the job.
A knowledge graph over the platform data
In developmentSo a question can traverse the real relationships between a job, an operation, a machine, a part and a supplier, and return the path it travelled, which is what makes the reasoning auditable rather than impressive.
Supplier email read against your open orders
Real, not scheduledAn add-in that reads inbound supplier mail with your SyteLine context attached, and surfaces the date change or the shortage before it becomes a stockout. Built, and paused.
A ladder, not a leap
Anybody can sell you the fifth rung. We are on the third, and the first two are delivered rather than described.
Six things this layer deliberately does not do
- AI for its own sake. Deterministic workflow where a rule will do, pattern recognition where the maths will do, and a language model only where the problem is actually language.
- Write back. The intelligence layer does not change your ERP. It reads, reasons and reports.
- Workforce and shift scheduling. A deliberate partner seam rather than something to build badly.
- Planning intelligence that belongs to your ERP vendor: material risk, inventory policy, supplier lead time, make against buy. That is their layer and they are good at it.
- Anything in human resources.
- Being a better chatbot. If the answer needs arithmetic, the arithmetic is not the chatbot part.
Both, not either
Their foundation is your ERP data, and it is good. The floor data is not in it unless something put it there.
We are not competing with your ERP vendor's AI. We are the layer that gives it something to see.
If the answer lives in your ERP, use theirs. If the answer lives between your people, your machines and your quality, that is ours.
Your ERP vendor is investing seriously in AI and it is worth turning on. It reasons about the record of work, which is what it has. This layer reasons about the work, because the platform underneath it created that data in the first place.
The other half of the moat: the model does not contain the business logic. It uses it at runtime. Copying that means rebuilding decades of SyteLine specific manufacturing knowledge, not buying a bigger model.
Ask the hard question first
Ask what happens when it is wrong, ask where the number came from, and ask to see the rows behind an answer. Those three questions separate a working layer from a demonstration, and we would rather you asked them of us than of somebody else.
Talk about your floorWhere this sits in the platform