Comparison
Why not just use ChatGPT?
For drafting, summarising, a one-off extraction from a pasted document or an ad hoc analysis, a general-purpose assistant is the right tool, and your team should use it. A workflow that runs every day across thousands of documents needs what sits around the model: access to your systems and write-back, validation against your own rules and master data, exceptions routed to a person, an audit trail and someone accountable for the result.
General-purpose assistant vs agent deployed in a workflow
| Aspect | General-purpose assistant | Agent deployed in a workflow |
|---|---|---|
| How the work starts | A person opens a conversation and brings the document or the question | The work arrives on its own (an order email, a freight document, a plan set) and the agent picks it up |
| Inputs | What one person provides, one request at a time | Every document from the mailbox, portal or folder the workflow already uses, in every format senders use |
| Systems of record | What the person copies in and out, unless someone builds and maintains a connection | Reads from and writes back into the ERP, TMS or estimating software, within the access the deployment defines |
| Validation | The person reads the answer and judges whether it is right | Checked against your business rules, price lists, rate tables and master data before anything is written |
| Exceptions | Caught if the person notices them | Flagged and routed to a named person, with the source document and the reason |
| Permissions and data boundaries | Depend on the assistant's account settings and on what each person chooses to paste | Designed per deployment: which systems, which records, which actions, inside the client's environment |
| Audit trail | The conversation, as far as anyone keeps it | Every input, extraction, check, decision and write recorded per document and traceable to its source |
| Volume and consistency | One request at a time, phrased differently by each person | The same instructions and checks applied to every document, thousands of times |
| When a format or system changes | Each person adapts how they ask | Monitored, and changed by the engineers who built the workflow |
| Who owns the outcome | The person who asked | The team that runs the system, with a person on your side signing off on exceptions |
| Best fit | Drafting, summarising, one-off extraction, ad hoc analysis | A recurring, high-volume workflow that ends in a write to a system of record |
Executive Summary
It is a fair question. General-purpose chat assistants are capable: they read documents, pull out fields, draft replies and reason through a problem in plain language. If someone on your team can paste a customer order into one and get the right lines back, why pay for anything else?
Because reading the order is the smallest part of the workflow. The work is in everything around it: getting every document in without someone pasting it, checking the result against your own rules and master data, writing it into the system of record, sending the cases that do not fit to the right person, keeping a record of what happened, and changing all of that when a supplier's format or your ERP changes.
That is the difference this page sets out. A general-purpose assistant is a tool a person uses. An agent deployed in a workflow is a system that runs inside your operation, with a person supervising it. Mirage builds the second kind; its agents use large language models among other components, and much of the engineering goes into the parts around the model.
Both have a place. The useful question is not which one is smarter, but which one fits the job in front of you.
Key Takeaways
- A general-purpose assistant is the right tool for drafting, summarising, one-off extraction and ad hoc analysis, with a person checking each answer.
- In an operational workflow the model is the smaller part: intake, validation, write-back, exception routing and audit are the rest.
- An agent deployed in a workflow checks its output against your own rules and master data before it writes anything.
- Cases that do not fit go to a person with the evidence, instead of depending on someone noticing.
- Someone has to own the result, and change the system when formats or systems change.
- Many teams will use both: an assistant for individual work, a deployed agent for recurring volume.
Key Differences
Where the work comes from
An assistant waits for a person to bring it something. An agent in a workflow is attached to where the work arrives: a shared mailbox, a customer portal, a plan room, a scanned document folder. Nobody has to remember to paste the hundredth order of the day, and none is skipped because the person who usually handles it is away.
Access to systems of record, and write-back
Reading a document is half the job. The other half is putting the result where it belongs: an order in the ERP, a shipment in the TMS, quantities in the estimating software's import format. That needs credentials, an integration that respects the system's own rules, and a plan for what happens when a write fails halfway. Mirage has connected more than 200 systems so far (ERP, TMS, WMS, maintenance, telematics, document systems), writing back into them.
Permissions and data boundaries
With an assistant, the boundaries are the account's settings and whatever each person decides to paste in. A deployed agent has its boundaries designed: which systems it can read, which records it can change, which actions need a person, and where it runs. Mirage's agents run inside the client's environment.
Validation against your own rules and master data
A capable model can read a customer reference correctly and still not know that this customer has its own price list, that an item was discontinued last month or that a lane is not served. Those checks live in your master data and in your team's habits. A deployed agent applies them before anything is written, on every document.
Exceptions go to a person, with the evidence
Some documents will not fit: an unknown product code, a quantity that does not match the drawing, a rate outside the agreement. In a conversation, the person spots it or does not. In a deployed workflow it is flagged and routed to the person who decides, with the source attached. At Biodéal, about 7% of orders are flagged for a human check, and the agent creates the orders in SAP Business One so nobody keys them in again.
Audit trail
When a customer disputes an order or an auditor asks why a value was entered, the answer has to be on record: which document, what was read, which rule applied, who approved the exception. A deployed system keeps that for every document, by design rather than by habit.
Volume, reliability and change
Thousands of documents a month need the same instructions and checks applied every time, a queue that absorbs peaks, and monitoring that notices when results drift or a sender changes a format. When the ERP is upgraded or a new document type appears, someone changes the workflow. That someone is part of what you are deciding on.
Ownership of the outcome
With an assistant, the person who asked owns the answer. With a deployed agent, a team answers for the system's output and its errors, and a person on your side signs off on what the agent cannot confirm. Mirage's forward-deployed engineers build the system and stay on it once it runs.
Advantages and Limitations
General-purpose assistant
Advantages
- Available now, with no project to start
- Strong at drafting, summarising and explaining
- Quick for a one-off extraction from a document someone pastes in
- One tool covers many different tasks
- Puts AI in the hands of every employee
Limitations
- Depends on a person to bring each input and check each answer
- Knows nothing of your master data unless someone builds that in
- Results vary with how each person phrases the request
- Write-back, per-document logging and exception routing have to be built around it
Agent deployed in a workflow
Advantages
- Picks up documents where they arrive, without anyone pasting them
- Checks every result against your rules and master data
- Writes into the ERP, TMS or estimating software
- Routes exceptions to a named person with the evidence
- Keeps an audit trail per document
- Someone is accountable for it and changes it when formats change
Limitations
- A project: the first workflow takes 4 to 6 weeks to reach production, longer across several systems
- Built for one workflow, not for any question an employee might ask
- Needs access to your systems and time from your team while the workflow is mapped
- Not worth it for low-volume or one-off work
Which one fits the job
Drafting a reply to a customer complaint
Assistant. One case, a person reads the draft, edits it and sends it. There is nothing to integrate and nothing to log beyond the email itself.
Summarising a long contract before a meeting
Assistant. A person reads the summary, checks the clauses that matter against the original and decides. Use it.
A one-off analysis of last quarter's claims export
Assistant. An analyst explores the data, asks follow-up questions and checks the numbers. A deployed system would be more than the job needs.
Entering every customer order received by email
Deployed agent. Orders arrive all day in every format, each needs the customer's own rules applied, and the result has to land in the ERP. At Biodéal, the median from the order email arriving to the order being ready in SAP Business One is 48 seconds.
Answering freight rate requests
Deployed agent, once the volume is there. Each request is priced from the company's own rates. At OFRET Groupe, more than 90% of standard freight RFQs are handled automatically, and a request outside the standard pattern goes to an operator before anything is sent.
Quantity takeoff across a bid's drawings
Deployed agent. Quantities have to be traced to the drawing they come from and delivered in the format the estimating team works in. At John Cockerill, piping takeoff takes 80 to 90% less time, with engineers reviewing the output instead of measuring.
Mirage Metrics
Platform
The governed environment that connects operational data, existing software, AI agents and the people who supervise them, cloud or on-premise.
See the pageForward-deployed engineering
How Mirage engineers observe the work, map its exceptions with your team, build the agent around them and run it in production.
See the pageUse both, for different jobs
The useful answer to the question is not no. Give your team a general-purpose assistant, agree what may and may not be pasted into it, and let people use it for the drafting, reading and analysis that fill their day.
Then look at the work that repeats: the documents that arrive every day, the checks that are always the same, the entries someone keys into a system by hand. That is where a deployed agent earns its place, because the value is not in reading one document well but in handling all of them correctly, writing the result back and showing a person only the cases that need one.
A simple test: if the task ends with a person reading an answer, an assistant is probably enough. If it ends with a record written into a system your business runs on, you need the rest of the workflow around the model.
What a deployed workflow adds around the model
Take one real workflow, customer orders received by email, and follow what happens around the language model once it runs in production.
- Intake. The agent watches the mailbox where orders arrive and picks up every attachment, whatever its format: PDF, scan, spreadsheet, free text in the body of the email.
- Reading. The model reads the document and extracts the customer, the items and the quantities. This is the step a general-purpose assistant already does well.
- Validation. Each line is checked against the customer record, the price list, the item master and the business rules your team applies today.
- Write-back. The order is created in the ERP the way a person would create it, so nobody keys it in again.
- Exceptions. An order the agent cannot confirm is flagged for a person, with the source document and the reason, rather than dropped or guessed.
- Audit. Every document, extraction, check, decision and write is logged and traceable to its source.
- Monitoring and change. Volume and exception rates are watched, and when a customer changes its order form or the ERP is upgraded, the engineers change the workflow.
- Ownership. A named team answers for the output, and a person on your side signs off on what the agent flags.
How Mirage uses language models
Mirage's agents use large language models among other components: document parsing, the rules mapped with your team, connectors to your systems, queues, logging, and the interface where your people review exceptions. Which model handles which step is an engineering decision made for each deployment.
Mirage does not sell an assistant or a login. The point of a deployment is not access to a better model, which your team may already have, but everything this page lists around it, built for one workflow inside your environment and run with your team.
ChatGPT is named on this page because it is the question buyers ask. The page describes general-purpose chat assistants as a category and makes no claim about any one product's features, plans, limits or data policies. Mirage Metrics is not affiliated with the makers of the assistants named here.
Further reading
The role behind this way of working, explained. See what a forward-deployed engineer is
Agents in production at Biodéal, OFRET Groupe, John Cockerill, Transwin and others, with their published results. See the case studies
FAQ
Related Comparisons
Build vs buy vs forward-deployed engineering
Build in-house, buy a SaaS product or bring in an outside team that builds and runs the system inside your environment: how to decide.
Read the comparisonAI Agents vs RPA
Why a fixed script breaks on varied documents and exceptions, and where it is still the right tool.
Read the comparisonOCR vs LLM
How character recognition and large language models differ in document automation workflows.
Read the comparisonMirage Metrics for operational workflows
Forward-deployed engineering
Mirage engineers observe how the work is done, map its exceptions with your team, build the agent around them and run it in production with you, starting from a proof of concept on your own documents one week after the first meeting.
See how forward deployment worksGet a custom automation plan
Tell us where your operations lose time. We send back a written plan: which workflows are worth automating, what a deployment looks like on your systems, and the shortest path to production. No commitment.
Prefer to talk it through? Book a call