AI Implementation Services
Most AI implementation puts a new tool on top of the old way of working, so the same broken process runs faster. We work differently: our forward deployed engineers work inside your business, map how the work really happens, redesign it, and build the automation into the CRM, accounting software and inboxes your team already uses.
Not technical, or new to the term? What is a forward deployed engineer?
Illustrative example: supplier invoices
Payments and exceptions still go to a person.
Almost every business knows this one. A customer fills in your website form or sends an email, and someone has to read it, research the company, check the CRM, write a proposal and remember to follow up. Here's that process before and after we redesign it.
Illustrative example: sales inquiry automation
Typical problems
Each step is tagged by who does it: simple rules, AI, or a person.
A person stays in control of the pricing and the final proposal.
| Before | After | |
|---|---|---|
| Hands-on time per inquiry | 20–30 minutes of manual work | 2–5 minutes of review |
| Tools opened by hand | Several: inbox, CRM, the company's website, past proposals | None: the details come to you |
| Response time | Sometimes hours | As soon as you approve the draft |
| CRM updates | Depend on who handled the inquiry | Automatic, every time |
| Proposal | Written from scratch or copied from the last one | A draft ready immediately, for you to check |
In most business processes the work itself takes minutes. The rest of the time is waiting: in inboxes, in queues, for a reply from another team, for a second approval.
An AI model that makes every step twice as fast saves a few minutes. Removing the waits saves days. That's why buying AI licenses rarely changes how a business performs: people like the tools, but the process underneath stays the same.
None of this is new. In 1990, Michael Hammer's Harvard Business Review article “Reengineering Work: Don't Automate, Obliterate” argued that companies were using computers to speed up old processes instead of redesigning them. AI hasn't changed the lesson: redesign the process first, then automate what's left.
Illustrative example: one supplier invoice
18 minutestouch time (hands-on work)
12 dayselapsed time, inbox to ready to pay
We deliver AI implementation through forward deployed engineers: software engineers who work inside your business instead of from a ticket queue. Ours own the whole job, from discovery and process redesign to production code, AI evaluation and the rollout with your team. New to the term? Read what a forward deployed engineer is.
| Typical AI implementation | Brainium forward deployed engineers | |
|---|---|---|
| Starts with | The list of tasks you hand over | How the work really happens, from interviews and system data |
| The process | Automated as it is today | Redesigned first, so unnecessary steps are removed before anything is built |
| Where it runs | A new tool or dashboard to log into | Inside your CRM, accounting software, inbox and Slack or Teams |
| Use of AI | One model for everything | Plain code where rules work, AI only where judgment is needed, chosen per step by testing |
| Proof | Delivered and handed over | Measured against a baseline taken before the build |
| After launch | Support tickets | Monitoring, evaluations rerun when models change, and the next process |
Every implementation starts with process mapping, from three sources, because each one alone gives you half the picture. Interviews without data miss the loops; data without interviews misses the reasons.
Why each step exists, who really decides, which steps are habit, and what happens when something goes wrong. None of it is written down. We do these ourselves, not with a form or an AI interviewer: the people whose work is changing have to trust whoever is changing it.
Read-only exports or API access to the CRM, accounting software, ticketing tool or inbox involved. Timestamps, edits, reassignments and reopened records show the real steps, the loops, and where work sits waiting.
Procedures, checklists, shared drives, spreadsheets and email templates. They're often out of date, but they show what the process was meant to be, and the gap is useful.
Two illustrative examples: sales inquiries at a services firm that uses HubSpot, and supplier invoices at a distributor that uses Xero. Pick one and step through the four stages of the map.
Illustrative example: sales inquiries at a 30-person services firm
What the sales process document says: five steps and no exceptions. Most written processes describe the happy path and nothing else.
What interviews with the sales team and a year of history in HubSpot and the shared inbox show: more steps, two loops, long waits, and two places where leads quietly fall through.
Each step goes into one of four buckets: remove it, run it as plain code, give it to an AI agent, or keep it as a human decision.
The new process. Nine of the ten steps run on their own. A person approves the pricing and the final proposal before anything goes out.
This is business process reengineering with AI in the toolbox, and it's where the savings are decided. It takes both software and AI experience, because the right answer depends on what code, models and people are each good at.
The step only exists because of a handoff, a re-key or an old workaround: renaming files, copying into spreadsheets, forwarding for visibility, chasing. The test: if it disappeared, would anything downstream break?
In the examples
If X then Y, with no judgment: matching within a tolerance, routing by amount, duplicate checks, scheduled syncs. Plain code is cheap, fast, auditable and never makes things up. If code can do it, we don't use AI.
In the examples
Judgment with history behind it. We use AI where there are hundreds or thousands of past decisions with known outcomes, the output is narrow (a field, a category, a route, a match) and a wrong answer is cheap to catch.
In the examples
High-stakes, irreversible or rare: approving payments, signing contracts, anything with no history to learn from. A person decides, and an agent prepares the case: the evidence, the history and a recommended action.
In the examples
We judge the difficulty of the action, not the difficulty of the thinking. A yes or no that a person can check in seconds is a good AI step. Five pages that a person has to reread is not.
No new platform and no migration first: replacing a CRM or ERP before automating is the slowest, most expensive route. Agents read and write through the APIs of the systems you already run, and people approve in Slack, Teams or email.
Webhooks, inbox rules and schedules start each run. We use the systems' APIs; screen automation is a last resort.
Every action is idempotent and logged with its inputs, so a failed run can be replayed without duplicates. Failures land in a queue your team can see.
Scoped service accounts with access only to the records each automation needs, read-only wherever possible.
The automation proposes and a person confirms until it meets the agreed accuracy on your real data. Then it acts on its own.
When two systems hold different versions of the same record, we agree up front which one wins and the automation follows that rule.
What came in, what was decided and why, what changed in which system, and who approved it.
Invoice needs your approval#finance-approvals · just now
Harbor Packaging Co. · INV-20417 · $6,480.00
Why it's here
Billed for 40 cases, but only 36 were received. It's also over the $5,000 limit.
Evidence attached
Suggested decision
Approve $5,832.00 for the 36 cases received, and request a credit note for the rest.
Approve suggested amountRejectSend to someone else
We're not tied to one AI provider. Most steps don't need the largest, most expensive model, and some are done better and cheaper by a smaller one. The only way to know is to test on your data.
For every AI step we build an evaluation set from your own past decisions and test several models against it. When a provider updates or retires a model, we rerun the evaluations before anything changes in production.
Your data stays protected throughout: we sign an NDA before we see any of it (plus a BAA or DPA where you need one), and we never use it to train AI models.
For every AI step:
Before any code is written we agree the numbers that matter and measure them from your own systems. After go-live we report against the same numbers.
Illustrative example: supplier invoices at a 40-person distributor
| Before | After | |
|---|---|---|
| Steps in the process | 13 | 6 |
| Elapsed time, invoice received to ready to pay | 12 days | Same day, or 2 days for exceptions |
| Touch time per invoice (average) | 18 minutes | About 2 minutes |
| Straight-through rate (no human touch) | 0%: every invoice is handled by hand | About 75%; people see only the exceptions |
| Loops that send work back | 3 | 1 |
| Invoices recoded at month end | 12% | Under 2% |
Your targets come from your own baseline, measured in the first phase. We usually start with the workflow that has the most handoffs, which isn't always the one with the most volume.
The same five phases for every client, whether you start with one process or a whole department. You approve each phase before the next one starts.
Phase 1 · Map
We interview the people who do the work, review the history in the systems involved and read the documents you already have. Then we measure where the time goes.
You get: The real process map, hands-on time compared with waiting time, the exceptions, and baseline numbers to measure against.
Phase 2 · Redesign
Every step is removed, handled by simple rules, given to AI or kept with a person. Together we agree the new process, who owns each step and what it should save.
You get: The redesigned process, the expected savings and a fixed quote. Nothing is built until you approve it.
Phase 3 · Build
We build the automation into the software your team already uses, test it on real data and run it alongside your team until it's reliable.
You get: Working automation, a short walkthrough for your team and simple documentation. Typically 2–6 weeks per process.
Phase 4 · Prove
After go-live we report the same numbers we measured in the first phase, so you can see exactly what changed.
You get: A before-and-after report you can share with your team.
Phase 5 · Improve
With Automation Care we monitor the automation, fix issues, update the AI as models change and start on the next process when you're ready.
You get: Automation that keeps working as your business changes.
The engineer who maps your process is the one who builds it, so nothing gets lost between a consultant and a developer.
We work in your CRM, accounting software, inbox and Slack or Teams. No new platform, and no migration first.
Every phase is quoted up front. If something isn't worth automating, we'll tell you.
We sign an NDA before we see any data (plus a BAA or DPA where you need one) and never use your data to train AI models.
What it costs
Your data is the one thing we can't get wrong. Here are the standards we meet and the agreements we sign before we work with it.
Independently audited security controls. The report is available on request, under NDA.
Certified information security management, covering how we protect your data.
HIPAA-compliant processes for protected health information, backed by a signed BAA.
GDPR- and UK GDPR-compliant processing, backed by a signed DPA.
Reports, certificates and agreement templates are available on request, under NDA. We'll confirm which ones apply to you during the assessment.
AI implementation services take AI from idea to working software inside your business. A provider maps how a process really runs, redesigns it, builds the automation into your existing systems, tests it on real data and measures the result. At Brainium, forward deployed engineers do all of it, from the first interview to go-live.
We start by mapping how the process really runs, from interviews with your team and the history in your systems. Then we sort every step: remove it, run it as plain code, give it to AI, or keep it with a person. We build the result inside the software you already use, run it alongside your team until it's reliable, and report the results against the baseline we measured at the start.
When a process crosses several people and systems, when nobody has the full picture of how it really runs, or when an earlier AI tool didn't change much. A forward deployed engineer is most useful where the process needs redesigning, not just a new tool. For one simple task, an off-the-shelf tool may be enough, and we'll tell you if it is.
An agency or freelancer usually automates the task you describe, the way it runs today. A forward deployed engineer first finds out how the process really runs, removes the steps that shouldn't exist, and only then builds. They also measure the result against a baseline taken before the build, so you can see what changed.
No. We build inside the systems you already run, through their APIs. When two systems disagree about the same record, we agree up front which one wins and the automation follows that rule. Migrating before automating is usually the slowest and most expensive route.
Whichever passes the tests for each step. We build an evaluation set from your own past decisions, test large, small and open-weight models against it, and use the cheapest one that meets the accuracy target. Many steps don't need AI at all and run as plain code.
We design for them. AI only handles steps whose output is narrow and easy to check, each AI step has an accuracy target that it must meet before go-live, rules validate the output, low-confidence results go to a person, and every action is logged. Anything that moves money or can't be undone stays with a person.
Only what each automation needs, read-only wherever possible, especially during discovery. We sign an NDA before we see any data (plus a BAA or DPA where you need one), and we never use your data to train AI models.
Yes. We're happy to work alongside your team. Every automation comes with a walkthrough and simple documentation, so your team knows what it does and how it's monitored.
A first look at one process usually takes days. Building the automation for one process typically takes 2–6 weeks from sign-off to go-live. A whole department is scoped individually.
Tell us where your team spends too much time. We'll map it, show you what can be removed and what can be automated, and tell you whether the numbers make sense.
Find What You Can AutomateFree assessment. No commitment. If something isn't worth automating, we'll say so.
