The engagement · how it runs
What the 90 days actually feel like.
Most vendors show you a four-box timeline and call it a process. This is the other thing you want to know before you commit: what happens in which week, who you deal with, how much of your time it eats, and what it looks like when something goes wrong.
Four phases, thirteen weeks, staged go-lives throughout — then a partnership that runs the whole thing afterwards.
- Build duration
- 90 days, fixed scope and timeline
- First system live
- Week 5–6, not day 90
- Your time
- Kickoff, then ~1 hour a week
- After day 90
- Month-to-month. 30 days' notice
01The part most people get wrong
Day 90 isn't launch day. Week 5 is.
A 90-day build where everything arrives at once on the last day is a 90-day bet. If the assumptions were wrong, you find out in month three with nothing to show and no time to fix it. So we don't build that way.
The usual shape
One launch, at the end.
Three months of build with nothing running. Every assumption from week one stays untested until week twelve. The first real feedback arrives the same day the invoice does, and if something was wrong, it was wrong for eleven weeks.
How we run it
Staged go-lives, one at a time.
Your highest-value agent ships first — usually week 5 or 6 — and starts paying for itself while the rest is still being written. Each system goes live as it passes testing. By day 90 you haven't been waiting on a launch; you've been watching it compound.
02The thirteen weeks
Every phase, what comes out of it, and what it asks of you.
Four phases. Each one lists what exists at the end of it and exactly what we need from you while it runs — including the phases where the honest answer is 'almost nothing'.
Map
Weeks 1–2Days 1–14We learn how your business actually runs.
Before anything gets built, we sit down and trace your business end to end — where leads come from, where they stall, what eats your team's hours, and which of those things is quietly costing the most. Then we design the AI team around what we found, in the order that pays for itself fastest.
What exists at the end
- A working session mapping your customer journey from first contact to paid
- The leak report: every place work or revenue is dropping, ranked by what it costs you
- An architecture for your build — which agents, in what order, and the outcome each one owns
- An ROI projection, so you know what it's worth before a line of code is written
What we need from you
One 90-minute kickoff working session, plus read-only access to the tools we're mapping.
Architect
Weeks 3–4Days 15–28We design the system before we build it.
Your senior architect selects the models, data stores, and integration points for your specific environment, and writes the build to spec. This is the phase most AI vendors skip, and skipping it is why most AI pilots never reach production: nobody decided in advance what the system was allowed to know, what it was allowed to touch, or what happened when it was unsure.
What exists at the end
- A technical specification with architecture diagrams and data-flow maps
- Model and data-store selection matched to your privacy and latency requirements
- Integration map against your existing CRM, phone, calendar, and accounting stack
- Escalation design: the confidence thresholds and human-approval checkpoints, agreed with you
What we need from you
About an hour to review and sign off on the spec. Nothing gets built until you have.
Build
Weeks 5–10Days 29–70Your agents go live one at a time, not all at the end.
The engineering phase, and the reason the sprint doesn't feel like waiting. Systems come online in sequence, tested against real scenarios from your own business, inside the tools you already use. Your first working agent is live weeks before day 90, so the build is paying for itself while the rest of it is still being written.
What exists at the end
- Your first agent live and handling real volume, typically in week 5 or 6
- Each subsequent system deployed as it passes testing, never in one big-bang launch
- Connected into your CRM, phone, email, calendar, and accounting — no rip-and-replace
- Weekly demos, and documentation written as we go so nothing is a black box
What we need from you
About an hour a week: a progress review and a go/no-go on each system before it ships.
Running by the end of this phase
Most of your AI team is running in production
Operate
Weeks 11–13Days 71–90The full build is live, and we stay on to run it.
The remaining systems go live, your team learns to work alongside them, and we move into operating mode. Day 90 is not the day things start working — plenty has been working for weeks. It is the day the build is complete and the partnership that keeps it working begins.
What exists at the end
- The full AI team live in production, with a documented runbook per agent
- Team training: recorded walkthroughs plus live sessions for the people using it daily
- Your live dashboard, showing hours and money the systems are saving
- Transition into the operating retainer — monitoring, tuning, and new builds on request
What we need from you
A training session per team that touches the systems, and sign-off on each go-live.
Running by the end of this phase
Everything is live, documented, monitored, and in operation
03Before the work starts
The first forty-eight hours, hour by hour.
How an engagement opens tells you most of what you need to know about how it will run. Here's ours, with the clock attached.
- 01Within 1 hour
Your kickoff is on the calendar
Not a scheduling email thread. You leave the strategy call with the working session already booked and the prep list in your inbox.
- 02Day 1
Your channel opens and your Specialist introduces themselves
A private channel with the people actually on your build — by name, with what each of them is responsible for. This is the line you use for the next several years.
- 03Day 1
Access requests, batched and explained
One list, sent once, with the reason each item is needed and the permission level we're asking for. Read-only wherever read-only is enough.
- 04Day 2
We start listening before we start asking
We go through your intake volume, your response times, and where conversations die — so the kickoff session is us confirming what we already found, not you explaining your business from scratch.
04The working rhythm
What a normal week looks like once it's running.
Not 'we'll be in touch'. A schedule you can put in your calendar and hold us to.
Any time
Your channel
A question, a problem, or a brand-new idea. Median first response is under a couple of hours in working time. No tickets, no queue, no hold music.
Weekly
A 30-minute progress review
What shipped, what's next, what we found, and anything that needs a decision from you. Recorded, so anyone on your side who missed it can catch up.
Continuously
Your live dashboard
Every agent's activity, the volume it handled, and the hours and money it saved — open it whenever you want rather than waiting for a report.
Quarterly
A strategy review
We step back from the build and look at the business: what changed, what's worth automating next, and what we'd stop doing if it were our money.
What the sprint asks of you, in total
- A 90-minute kickoff working session, then about an hour a week
- Access to the tools we're connecting — we handle every piece of the wiring
- Sign-off at each go-live: nothing reaches your customers without your approval
- One person on your side who can answer questions about how the business really works
Your team keeps doing their jobs throughout. The build happens around them, not on top of them — and no system reaches a real customer without your sign-off first.
05Division of labour
Our side of the line, and yours.
The most expensive misunderstanding in this industry is a client who thought they were buying an outcome and discovers in week five they bought software to staff. So here it is in two columns, before you sign anything.
Echelon owns this
- Every line of the build, the integration work, and the testing
- Choosing the models and the architecture, and defending those choices to you
- Running the systems in production: monitoring, alerting, and the 3am failures
- Keeping agents current as the models change underneath them
- Writing the documentation and the runbook for every agent
- Telling you when something you asked for isn't worth building
You own this
- One kickoff working session, then roughly an hour a week
- Access to the tools we're connecting — we do all the wiring
- Straight answers about how the business really runs, including the ugly parts
- A go/no-go before each system touches a real customer
- Someone who can settle a judgment call when we hit one
- Telling us when an agent gets something wrong, so it stops getting it wrong
06What could go wrong
Four ways these builds go sideways — and what we do about each.
Every one of these has happened, to us or to someone we've had to clean up after. A vendor who can't name their own failure modes hasn't shipped enough to have any.
Your data is messier than anyone thought
The honest version
It usually is. Duplicate contacts, four spellings of the same company, a decade of notes in a field meant for phone numbers.
How it's handled
We find this in Map, not in week 8, because we go through the real records before we design anything. If cleanup is needed we scope it and price it separately instead of quietly eating the timeline.
An agent gets something wrong in front of a customer
The honest version
Over enough volume, this happens. Anyone who promises otherwise is either new or not telling you the truth.
How it's handled
It's designed for. Confidence thresholds mean an unsure agent escalates to a human with full context rather than improvising. Anything that moves money or signs your name waits for a human tap. And every step is traced, so we can tell you exactly why it did what it did.
Your team quietly refuses to use it
The honest version
The most common way an AI deployment dies. The system works fine; people route around it and keep doing it the old way.
How it's handled
We build into the tools they already open, so there's no new thing to log into, and we train the people doing the work rather than only the person who signed. If usage drops, it shows on the dashboard and we come ask why.
A tool you depend on changes or breaks
The honest version
APIs deprecate, vendors get acquired, a model provider changes behaviour on a Tuesday with no notice.
How it's handled
Agents degrade gracefully instead of failing silently: if a tool is unreachable the agent says so and queues the work rather than dropping it. Alerting comes to us first, and staying ahead of this is a large part of what the retainer is actually for.
07Day 90 and after
The build ends. The partnership doesn't.
Most AI projects die quietly in month four: the person who championed it moves on, nothing gets tuned, a model changes underneath it, and eventually people stop trusting it. The retainer exists specifically to prevent that.
We watch it so you don't have to
Alerting is wired to our on-call, not to your inbox. Most issues are found and fixed before anyone on your side notices there was one.
We keep it current
Models change, tools deprecate, your pricing moves. Agents get retuned as that happens, so month six performs better than day 90 rather than worse.
New ideas just go in the channel
You don't scope a new project to add an agent. You describe what you want in the chat and it gets built, prioritised in the weekly review.
You can leave whenever
Month-to-month, thirty days' notice, no penalty. You keep every outcome, all your data, and the documentation for how the operation runs.
08Questions about the process
The things people actually ask on the second call.
The 90-Day Sprint
Start the map
The first hour is the map. It's also free.
Book a 30-minute call and we'll trace where your business is leaking work, tell you which agent would earn its keep first, and give you the honest answer on whether this is worth doing at all.
Prefer to talk now? Call or text (805) 285-2713
- No commitment
- 30-minute call
- We'll tell you if it's not a fit