All Comparisons/Comparison

Comparison

Build vs buy AI agents for operations

Buy a product when the workflow is standard and lives in one system. Build in-house when the workflow is how you compete and you have engineers to run it for years. Hire a systems integrator when the need is stable, fully specified and your IT department can run the result. Bring in a forward-deployed team when the workflow is specific to your documents and rules, spans several systems, carries real volume, and nobody in house can build and run it. Mirage is that fourth kind: at Biodéal, its agent turns customer emails into orders in SAP Business One with a median of 48 seconds.

Published Last updated

Short answer

For an IT buyer, the choice turns less on the model than on five questions: who owns the code, what your security team has to review, who writes and maintains the integrations, what it costs to run after go-live, and what you keep if you stop. The decision table further down answers them for each option.

The options at a glance

OptionFits whenWatch for
Build in-house (your engineers build and run it)The workflow is core to how you compete, you have engineers with machine learning and integration experience for the long run, and you want to own the code.Time to production follows hiring and priorities. Every new format, model change and system upgrade lands on your team.
Buy a SaaS product (a vendor's product, configured by your team)The workflow is standard, stable and lives mostly in one system, and a product already does it the way you work.Your process adapts to the product. What is outside its scope stays manual, and its exceptions come back to your team.
Hire a systems integrator or consultancy (they build to a specification and hand over)The need is stable and can be written down in full, your IT department can run the result after acceptance, and you want to own the code without hiring the team that writes it.The engagement ends at acceptance; rules found after go-live, model changes and new formats land on your team. Code ownership is whatever the contract says.
Forward-deployed team (an outside team builds and runs it inside your environment)The workflow is specific to your documents, rules and several systems, the volume is real, and you have no engineers to build and run it.Changes go through an outside team, and scoping takes your people's time up front. Not justified for a standard workflow or low volume.
Keep it manualVolume is low, the work is occasional, or the process still changes too fast to automate.It does not grow with volume, and the knowledge stays with the people who do it.

Executive Summary

Automating an operational workflow with AI is usually framed as build versus buy. Building gives you a system made for your workflow, if you have the team. Buying gives you a product that works quickly, if your workflow looks like the one it was designed for. Between the two sit two outside options: a systems integrator that builds to your specification and hands over, and a forward-deployed team that builds inside your environment and keeps running the system.

Mirage is a forward-deployed team, which is not the right answer for every workflow; this page says when it is not.

Six questions that decide it

How specific is the workflow?

If many companies run it the same way, buy. If it depends on your own formats, customer rules, rate tables or drawing conventions, a product covers the common part only. Specific and core points to building; specific and not core, to an outside team.

How many systems does it touch?

One system with a product built around it: buy. Several, with writes into more than one, and the integration becomes the project: your engineers, an integrator if the interfaces are stable, or a forward-deployed team as part of the deployment.

Do you have engineers who can build and run it?

Building needs people who work with language models, document parsing and integrations, and who stay after launch. If IT can run a system but not build one, an integrator can build it and hand it over. If nobody can do either, building starts with hiring.

Who owns the exceptions?

An unknown product code, a rate outside the agreement. A product sends them back to your team; an integrator routes the ones its specification names, and later ones are yours. With a forward-deployed team the routing is mapped with your team first, and a person on your side signs off on each.

How soon does it need to be in production?

A configured product is usually the quickest start; a build depends on staffing, an integrator on the specification. Mirage runs a proof of concept on your documents one week after the first meeting and puts the first workflow in production in 4 to 6 weeks.

How often will it change?

Templates, rules, systems and model versions all change. A product follows its vendor's roadmap, an in-house system your team's time, an integrator's system a change request, a forward-deployed team's system its running.

Build and buy: advantages and limits

Build in-house

Advantages

  • Built exactly around your workflow and systems
  • You own the code and the roadmap

Limitations

  • Needs language model and integration engineers, kept on after launch
  • Maintenance and exceptions are permanent work, not a project

Buy a SaaS product

Advantages

  • The quickest start for a standard workflow
  • Maintained by the vendor, with no team to staff

Limitations

  • Your process adapts to the product, not the reverse
  • What is outside its design stays manual

Five workflows, five answers

A standard workflow inside one system

Supplier invoices in a common layout, posted into one accounting system, with a product built for exactly that. Buy it.

A fixed interface your IT team will run

A nightly export from the ERP to a carrier portal, specified once, at a company whose IT already runs its integrations. Hire an integrator, have the code assigned to you, price a maintenance option up front.

Customer orders by email into the ERP

Forward-deployed. At Biodéal, Mirage's agent creates each emailed order in SAP Business One, with a median of 48 seconds from email to order ready and about 7% of orders flagged for a human check.

Takeoff on an engineering group's drawings

Forward-deployed. At John Cockerill, piping takeoff takes 80 to 90% less time, with engineers reviewing the output.

Low-volume or occasional work

Keep it manual. A handful of documents a week, or a process still being redesigned, justifies none of the four.

Why build versus buy is not the whole choice

The hard part is rarely the model. It is fitting a system to a workflow that exists only in your documents, systems and habits, and keeping it fitted. A vendor provides no engineers for that, an integrator until acceptance, a forward-deployed team for as long as the system runs.

The IT buyer's decision table

What an IT or procurement team asks before signing, for the four ways to put an AI agent into an operation. These are the usual arrangements; a contract can move any of them.

Build, buy, systems integrator and forward-deployed team compared on code ownership, data and keys, security review, integration, run cost, change and exit
What IT asksBuild in-houseBuy a productSystems integratorForward-deployed team (Mirage)
Who owns the code?You do, for code your employees write at work. Contractors in the team need a signed assignment.The vendor. You hold a licence to use the product, and your configuration lives inside it.Whoever the contract names. Without a signed written assignment, the copyright in custom code usually stays with the integrator.The platform stays the provider's, licensed through the subscription. Who owns the rules, connectors and configuration built for you: ask before the pilot.
Where do the data, logs and keys sit?In your environment, under your controls.In the vendor's environment unless it offers your cloud or your servers. The model provider it calls is one more processor to approve.Usually in your environment, since the system is handed over to your IT department.Your infrastructure, your cloud account or hosted by Mirage. In the first two, data stays in your environment and the encryption keys are yours.
What does the security review cover?Your own design review: access, secrets, prompt injection, what the agent may write, how its output is validated.The vendor's questionnaire and attestations, its sub-processors, where data is stored, and what its agent is allowed to write into your systems.The code at acceptance and the integrator's access during the project. After handover, patching is your team's.The provider's access (at Mirage, limited to what the deployment needs, and you can remove it), model provider retention, write permissions per workflow, the audit log.
Who writes and maintains the integrations?Your engineers, for every connector, and again at each upgrade of the ERP, TMS or estimating tool.The connectors the product ships with. Beyond them, its API, and your team or an integrator.The integrator, to the specification. Then a maintenance contract, or your IT department.The deployment team, for as long as it runs. Mirage has connected more than 200 systems so far.
What does it cost to run after go-live?The people who keep it running, model usage and hosting, for as long as it exists.The subscription, plus the integration and exception work the product leaves to your team.The maintenance contract or your IT team's time. Every change after acceptance is a new change request.A subscription priced by volume, not per user: at Mirage, from $1,000 a month after a launch from $10,000.
Who handles a new format, rule or model?Your backlog.The vendor's roadmap.A change request, priced and scheduled.The team running it, as part of running it.
What do you keep if you stop?Everything, and the knowledge, as long as the people who built it stay.Your data, as the contract and data protection law require. The configuration usually stays behind.The code and documentation the contract assigns to you, and a system your team now has to run.At Mirage, you can stop after the fixed-price pilot with the output and the comparison in hand. For a running deployment, write the handback into the contract: data, logs, documented rules.

On code ownership: in the United States, custom software commissioned from an outside firm is generally not among the nine kinds of commissioned work that can be a work made for hire (17 U.S.C. § 101), and a transfer of copyright is valid only in a signed writing (§ 204(a)). In the UK, a computer program is a literary work (CDPA 1988, s. 3(1)(b)), its author is the first owner, and an employer owns it only when an employee wrote it in the course of employment (s. 11(1) and (2)). This is a summary, not legal advice: your counsel reads the contract.

On data: under the GDPR, a processor may not engage another processor without your prior written authorisation, and at the end of the service it deletes or returns the personal data, at your choice (art. 28(2) and 28(3)(g)). In the EU, the Data Act gives customers of cloud and other data processing services the right to switch provider, and removes switching charges, data egress included, from 12 January 2027.

Mirage's answers come from its enterprise and security pages: your data never trains a model, and the model providers a deployment calls keep nothing once a request is answered.

The integrator or consultancy option: what it delivers and what it leaves behind

A strategy consultancy delivers a diagnosis and recommendations that someone else implements. A systems integrator, or IT services firm, builds to a specification, tests against it and hands the system over at acceptance. The integrator is the one that competes with building and buying.

It leaves behind code, documentation and, if you bought one, a maintenance contract. With an AI agent it also leaves everything that is not code: the prompts, the documents the system was tested on, the model version, the exception queue and who clears it. After acceptance those are your IT department's to run, and so are the rules that only appear in production. The test is simple: if your IT team already runs integrations and can own a system that changes when suppliers change their documents, an integrator gets you a system you own; if not, the system drifts away from the work, one unhandled format at a time. What to write into the contract:

  • Code ownership. A written assignment of the custom code, with the integrator's pre-existing tools named and licensed to you.
  • Handover. Source code, prompts and configuration, the test documents with expected results, the model version, and a runbook for the exception queue and a model change.
  • Access and knowledge. The integrator's access removed at handover; named people on your side trained before acceptance.
  • After acceptance. A maintenance option and the price of a change request, agreed in the original contract.

What the security review should cover for an AI agent

Whoever builds it, an agent that reads documents and writes into your systems adds risks a classic integration does not have. The OWASP Top 10 for LLM Applications (2025) names them; four matter most here, and the review needs an answer to each, whoever builds the system.

  • Prompt injection (LLM01). A document or an email can carry instructions. What stops a supplier's PDF from telling the agent to change a price or skip a check?
  • Sensitive information disclosure (LLM02). What data reaches which model provider, under what retention terms? That provider is a processor you approve.
  • Improper output handling (LLM05). Is every value the agent produces validated against your master data and rules before it is written anywhere?
  • Excessive agency (LLM06). What can the agent write, into which system, under which account, and which actions need a person? Then the audit trail: at Mirage, every agent action is traceable to the input and the rule that produced it, and the logs stay under your control.

When not to buy

Buying is the right default for a standard workflow, and the wrong one in five situations that are easy to spot before the contract.

  • The workflow is how you compete. A pricing model, a way of quoting or a method of estimating that wins you work should not live in a product your competitors can buy too.
  • It cannot write back. If the product's output has to be re-keyed into your ERP, TMS or estimating tool, you have bought a second screen, not automation.
  • Your data cannot leave your environment. If the product runs only in the vendor's cloud and your contracts or regulator say the data stays with you, the review will stop it.
  • Your volume is in the exceptions. The product removes the easy cases and leaves the work.
  • There is no clean exit. If your data, logs and configuration cannot come out in a usable form, the switching cost grows every month.

Forward deployment: advantages and limits

An outside engineering team builds the system inside your environment, around your workflow, and runs it with you once it is live. At Mirage the method has five steps: Observe, Map, Build, Run, Expand.

How forward deployment differs from consulting, integrators, software vendors, presales and staff augmentation, model by model, is in the forward deployed engineering guide. See the comparison of delivery models

  • Run, not handed over. The engineers who build the system, write-back included, stay on it in production.
  • Limits. Changes go through an outside team, so settle code, data and configuration ownership in the contract; mapping the work takes hours from the people who do it.

How do you score a workflow for build, buy, integrator or deploy?

Score each line 0, 1 or 2, add them up and read the total against the bands. It is a prompt for the steering meeting, not a verdict: one line that matters to you can outweigh the total.

  • Specificity (0 to 2). 0: many companies run it the same way. 1: standard with a few of your own rules. 2: depends on your document formats, customer rules, rate tables or drawing conventions.
  • Systems touched (0 to 2). 0: one system. 1: two, with a scripted hand-off. 2: three or more, with writes into more than one.
  • In-house engineers (0 to 2). 0: a team with model and integration experience, with time. 1: an IT team that can run a system but not build this one. 2: nobody who could build or run it.
  • Exception load (0 to 2). 0: rare, handled by a person. 1: regular, with known causes. 2: exceptions are the normal case.
  • Urgency (0 to 2). 0: next year is fine. 1: this quarter. 2: the backlog is costing bids, orders or shipments now.
  • Rate of change (0 to 2). 0: stable for a year or more. 1: a new supplier format or rule every few months. 2: formats, rules or systems change monthly.
  • Reading the total. 0 to 4 with low specificity: buy, or keep it manual. 5 to 8 with engineers at 0: build. Engineers at 1 with change and exceptions at 0: an integrator. 7 to 12 with engineers at 1 or 2 and specificity at 2: a forward-deployed team. Anything else turns on the line you care about most.
  • A worked example, invented figures. Customer orders by email at a distributor: specificity 2 (each customer's own format), systems 2 (mailbox, ERP, a portal), engineers 2 (an IT team of three running the ERP), exceptions 2, urgency 1, change 1. Total 10 with specificity 2: the forward-deployed band. The same distributor's supplier invoices in one common layout would score about 3: buy.

What the research says

In MIT NANDA's The GenAI Divide: State of AI in Business 2025 (January to June 2025: 52 interviews, 153 leaders surveyed, more than 300 public AI initiatives), pilots built with external partners reached deployment about 67% of the time, against about 33% for internal builds (p. 19). The authors warn that the correlation does not prove causation, and note that the best buyers held AI vendors to benchmarks closer to consulting firms or outsourcers than to software (p. 20).

Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls (press release, 25 June 2025). Each of the three is a question in the decision table above: run cost, the workflow chosen, and the security review.

On the cost of building, Sculley and colleagues at Google showed that machine learning systems incur large ongoing maintenance costs, from data dependencies, configuration and changes in the outside world (Hidden Technical Debt in Machine Learning Systems, NIPS 2015). An agent that reads supplier documents meets the outside world every day.

When Mirage is not the right choice

Said plainly, so a first meeting is not spent finding out.

  • A standard workflow in one system. If a product already does it the way you work, buy the product.
  • A team that wants to own the system. If you have the engineers and owning the code matters, build it. Mirage sells no self-serve software.
  • A fixed need your IT team will run. An integrator that hands over the code is the better fit.
  • No document or exception volume. A few documents a week do not justify a deployment.

How a Mirage engagement runs

A first meeting to scope the workflow. A proof of concept on your own documents one week later. A fixed-price pilot on site, which you can stop after. From $10,000 to launch, then an annual subscription from $1,000 a month, priced by volume.

The first workflow is in production in 4 to 6 weeks. Workflows that span several systems can take longer and are scoped at the first meeting. Mirage has more than 35 deployments in production, for more than 30 clients in six countries.

Further reading

Where Mirage runs, who holds the data and the keys, and what its access is. See security and data handling

What an agent may do without a person, and what it logs. See approvals and audit

Deployments at Biodéal, OFRET Groupe, John Cockerill, Transwin and others, with their published results. See the case studies

FAQ

  • Buy a standard workflow in one system that a product already does. Build a core workflow if you have engineers to run it. Hire an integrator for a stable, specified need your IT can run. Bring in a forward-deployed team for a specific, multi-system workflow nobody in house can build or run.

  • Whoever the contract names. In the US, commissioned software is generally not a work made for hire and a copyright transfer needs a signed writing (17 U.S.C. 101 and 204(a)); in the UK the author owns it unless an employee wrote it at work (CDPA 1988, s. 11). Without a written assignment, the integrator usually keeps it. Have your counsel read the clause.

  • Who owns the code and configuration, where data, logs and keys sit, which model provider it calls on what retention terms, what the agent may write and which actions need a person, how injected instructions in a document are stopped, and what you get back when you stop.

  • When the workflow is how you compete, when it cannot write back into your system of record, when your data must stay in your environment and the product runs only in the vendor's cloud, when the work is in the exceptions, and when there is no clean exit.

  • There is no honest single figure for building, buying or hiring an integrator: a product is usually the quickest start, a build depends on hiring, an integrator on the specification. Mirage runs a proof of concept one week after the first meeting and puts the first workflow in production in 4 to 6 weeks.

  • Building costs engineering time, for the first version and for as long as the system runs. Buying costs a subscription plus the integration and exception work the product leaves to you. Mirage runs a fixed-price pilot that you can stop after: from $10,000 to launch, then an annual subscription from $1,000 a month, priced by volume.

  • In house, everything, while the people stay. With a product, your data, rarely the configuration. With an integrator, what the contract assigns you. With Mirage, you can stop after the pilot with the output and the comparison in hand; for a running deployment, write the handback into the contract.

Forward deployment at Mirage

Forward-deployed engineering

Engineers observe how the work is done, map its exceptions with your team, build the system around them and run it in production with you, then move to the next workflow once the first is trusted.

See how forward deployment works

Show us the workflow that costs you the most time

Leave your work email, then pick a time for a 30-minute call. We look at one of your workflows together, and you leave with a written plan to automate it. No commitment.

We use your email to reply to you and, now and then, to tell you about Mirage. You can opt out at any time. Privacy policy