From first look to first decisions.
An implementation service for small and mid-size companies without an AI team. We build fast, iterate fast, bring the know-how and train your people to run it. Then we stay, and keep tuning each agent on its numbers. Each stage moves one kind of work up one autonomy level.
Stage 0: Observe
We map the work before we touch it.
- Observer agents map key roles and repeated tasks
- Read-only access, with every read logged
- A ranked automation plan: signal, owner, value
- A sponsor
- Read-only CRM and shared-mailbox access
- 3 to 5 interviews with the busiest roles
Our proof. Our observers ran daily from July 2026. We started with 25 automation ideas and kept the 8 that mattered.
Stage 1: Draft
Agents do the prep. Your people approve it.
- A daily digest to named leaders, with the action written out
- Rules tested on your own past cases first
- A filter that keeps junk out
- Named recipients
- Feedback on the first 10 digests
- Decisions on what is worth flagging
Our proof. On its first run, our sales agent surfaced 607 open items.
Stage 2: Act reversibly
Internal actions you can undo.
- Agents open items, post internally, ask owners and log notes
- A list of allowed actions and an off switch
- Customer email, money and contracts stay with your people
- The allowed actions and spaces
- An owner per action type
- Someone who holds the off switch
Our proof. Every write our agents make to our operations system goes through its own identity, with limits, an audit trail and undo.
Stage 3: Decide
Low-risk calls, inside your limits.
- The agent makes one agreed kind of decision and records why
- A person reviews the log and can reverse
- A limit for each agent, raised only by a named approver
- Which decisions to hand over
- The approver
- The review cadence
Our proof. Every step up in our own agents' freedom was approved in writing by a named person.
Your own install, configured for you.
Our proven stack, set up for your roles, your systems and your keys. Nothing is shared with another client.
- A dedicated install per client
- Your own keys and credentials, isolated in your install
- Google Workspace connectors from day one
- Microsoft 365 next, and other connectors on demand
- Results read from each agent's own profit and loss
Run continuously
After go-live we keep watching and adjusting.
- Each agent's profit and loss, health and open issues, live
- Regular reviews with one of your people
- Rules and autonomy adjusted on the numbers
- Independent checks on what agents claim
- A typical change: raise the bar for the digest, so staff act on more of it
Ninety days, six months, twelve months.
First 90 days
Observe, then draft.
- Stage 0: map roles, write the plan
- First agent drafting, its numbers live
- Tuning on the numbers from day one
- Lattice switched on
By 6 months
Act reversibly.
- Stage 2: undoable actions with owners
- Independent audit and inventory in place
- Second and third agents
- Autonomy raised where the numbers support it
By 12 months
Decide, with oversight.
- Stage 3 for low-risk decisions
- More apps as they are ready
- Savings reviewed agent by agent
- An embedded managed service
Typical timeline. We set it with each client.
You will recognize the signs.
Who
- Small and mid-size companies, any industry
- Quotes, orders, work and invoices across several systems
- A busy team, with no one dedicated to AI
- Buyer: owner, CEO, COO or CFO
Signals
- Customers report the stalled quote first
- Two systems disagree on one fact
- Mornings lost to reading mail
- A past AI pilot made a demo, not a routine
- Experienced staff are retiring
Triggers
- A late quote cost a deal
- Unbilled work found at quarter end
- A key person leaving
- A new CRM or ERP
- The board asks about AI controls
Not a fit: you want agents emailing customers on their own from day one, there is no owner for the outcome, or there is no system of record for quotes or orders.
Energy and field services
Field paperwork drafted, quotes chased, compliance records kept, and your experts' know-how written down before they retire.
Financial institutions
Loan and account paperwork chased, onboarding checks prepared for review, and audit evidence gathered as work happens. Money movement and credit decisions stay with your people.
Healthcare offices
Scheduling, supply orders, credentialing and vendor paperwork for the business side of the office. Patient records and clinical decisions stay out of scope.
Straight answers.
Whose software is this?
Ours. The agents, the apps and the shared memory are CrystalAgent products. The agents use leading commercial AI models underneath.
Will an agent email my customers on its own?
Never. A message to a customer is a draft until a named person approves it, and the approval and the send are both logged. The same goes for money, contracts and credentials. The agent prepares the finished work and hands it over ready to approve.
Will it make things up?
Facts come from a source or are marked to confirm. Mind has a could-not-verify layer that shows every gap, and the independent auditor checks claims against the record.
Our last AI pilot produced a demo and nothing else. Why is this different?
We replay your real past cases first and move one level at a time. Each agent has a named owner and a profit and loss you can read from week one.
What about security?
Read-only by default. You get a dedicated install with your own keys and credentials. Your mail and files stay in that install: the search index the agents use lives there, under your keys, and you can export or delete it at any time. We keep no copies on our side. Access is removed in one step, and every action and approval is logged.
Will my staff drown in alerts?
Each alert has one owner, the digest is capped and junk is filtered. Technical problems go to an operations feed and never into a business digest.
What if an agent goes wrong?
You control an off switch. Agents change only what you approve and can undo it. Two corrections inside 14 days and a capability steps back one level.
Will auditors accept it?
Every action is logged with its reasoning, and Assay independently checks what agents claim. The audit reports to the owner, not to the people who built the agent.
How long until we see value?
Work that your people use shows on the agent's profit and loss from the first week. A first agent is usually drafting within the first 90 days.
Make your work crystal clear.
Request a 45-minute call. We pick one busy role in your company and show what its agent would save.

