Product 02
Agents that do the work, not ones that summarise it. They read the thread, parse the purchase order, open every attachment, match the product, apply your price tables and write the reply — then ask a human when they aren't sure, and remember the answer.
Underneath all of them is one model of how this industry actually works. We built a full ERP to learn it. Where your data lives is a deployment decision, not a limit on what they understand.
The fastest route. Price tables, catalogue, equipment and customer history already sit where the agents expect them, so there is nothing to connect and nothing to map.
You have an ERP, it works, and replacing it is not this year's project. Then it stays. We wire the brain into the stack you have and become the layer that understands the work — your systems keep being the system of record.
No two stacks are the same, so no two builds are. This one starts with a scoping engagement, not a signup.
The working loop
This is the whole job, and the agents own all of it. Every step is something a person in your office is doing right now, between other things, at the wrong time of day.
Quantities, sizes, colourways, decoration methods, placements, in-hands date, ship-to address, billing terms. Pulled out as structured fields, with the sentence each one came from.
The same applies whatever channel the job arrives on — an email thread, a purchase order document, a portal submission or a phone call. One request format, one queue.
Vector files, PDFs, images, a phone photo of a sample garment. It counts distinct colours, estimates stitches at the requested sew width, checks whether the resolution will hold up in production, and tells you when the customer's "two-colour logo" is really five.
Free text becomes a catalogue product, a colourway, a size run and your preferred vendor for that item. If two products are plausible, it says so rather than picking one and hoping.
Quantity break, colour count, stitch count, number of locations, set-up and screen charges, and the minimum that applies to the press that can actually run the job.
You decide what a complete request looks like — by default, every field on your order and purchase-order forms except mockups. Anything missing gets asked for, politely, in your voice, on a schedule you set.
At 100% complete, it can draft the quote, create it in the system, and move the deal to the right stage. Or it can stop and wait for a human. That's a dial, not a hard-coded behaviour.
Purchase orders
Contract work doesn't come as a friendly email. It comes as a purchase order — a PDF, a spreadsheet, a portal export, a scan of a printout — and somebody in your office retypes it into the system, line by line, at the busiest hour of the day.
The agents read it instead. Line items, SKUs, colourways, the size matrix, in-hands date, ship-to, billing terms and PO number come out as structured fields, matched against your catalogue and that account's pricing. Order entry becomes a review step: nine lines confirmed, one flagged, thirty seconds of your attention instead of twenty minutes of typing.
Voice
The same agent, on the same brain, taking calls. It knows your products, your pricing and the customer on the line — so a call at 6:40pm becomes a request in the system instead of a voicemail somebody listens to tomorrow.
Nobody wants a robot answering every call. So the rules are yours: which hours it takes, whether it only catches overflow when the line is busy, which numbers it never answers, and who it hands off to.
Every call is transcribed and filed against the customer and the order, the same as an email thread. If the caller describes a job, it comes out as the same structured request — quantities, sizes, decoration, dates — and joins the same queue, with the same chase-what's-missing behaviour.
“We need about 150 hoodies for a staff event, left chest logo, need them the week of the 22nd.”
Outside answering hours the same call would have gone to voicemail and been actioned on Wednesday.
Also on the agent
Reading the enquiry is the visible half. These are the jobs it does once the order is real.
Places the logo on the real garment at the real print size, using the supplier's own size chart. When somebody nudges it by hand, that correction becomes how it places the next one.
Puts the embellishment stage on a press that can actually run it, on a day with capacity, respecting colour limits and minimums. Override it whenever you like.
Drafts the follow-up when a customer goes quiet, reminds an approver at 48 hours, asks for a production sample photo before a long run ships, and builds the quote the moment a request hits 100%.
Control
Nobody sensible hands a new hire the keys on day one, and an agent is no different. InkIQ runs at one of three levels, set per account and changeable at any time.
Every outbound action is logged with what it did, why, and what it read to decide. Confirmation gates sit in front of anything that reaches a customer.
Drafted follow-up: “Happy to get this priced today — could you confirm the size breakdown and the delivery address? The artwork you sent is print ready.”
Learning
Not by training on your data in a shared model. By writing down what happened in your account, for your account only.
In autonomous mode it messages the owner with the specific question and the context behind it. The answer becomes a rule it applies from then on.
Move a logo, rewrite a draft, change a price on a quote it built — the difference between what it proposed and what shipped is captured automatically.
Rules learned in your account apply only in your account. The work improves the system's judgement over time, but your artwork, pricing and customer data are never shown to another company or reproduced outside your account.
The obvious question
It will. Any shop owner evaluating this is really asking one thing: what does a mistake cost me, and how would I even know? Straight answers.
Only if you let it. Out of the box the agent prepares and a person sends. Even at the most autonomous setting, anything that reaches a customer can sit behind a confirmation, and pricing isn't invented — it's your table, applied. If the table doesn't cover a case, it escalates rather than improvising a number.
Colour counts and stitch estimates are shown as findings you can see and change, next to the file they came from, before anything is quoted. Where it isn't confident — a gradient that may or may not be halftones, small text that may not hold — it says so in the finding instead of committing to a number. It also tells you when a customer's own description disagrees with the file, which catches more errors than it causes.
When two catalogue products plausibly match a description, it flags both instead of choosing. "Two products match this description" is a normal outcome, not a failure — the aim is a system that hands you the ambiguous 10% rather than one that quietly guesses on all of it.
Every action is logged with what it did, when, on whose behalf, and what it read to decide. If a customer asks why they received something, you can answer precisely rather than reconstructing it. Requests that escalate appear as work waiting for a human, not as silent failures.
Corrections are captured. When someone rewrites a draft, moves a logo or changes a price on something the agent built, the difference between what it proposed and what shipped is recorded and applied next time — in your account only. Answers you give it when it escalates become rules it follows from then on.
It asks, and it says what it's unsure about rather than producing a confident-sounding guess. In autonomous mode that's a message to the owner with the specific question and the context behind it. A system that stops is worth considerably more than one that's right most of the time and never tells you which times.
The assistant
The same agents sit in a panel beside whatever you're working on. They know how the software works and can query your actual data, so the answers are about your business rather than a help article.
“How many black heavyweight tees in size small did we embroider with a left-chest logo last quarter, and what was the average margin?”
It can also do things: create a task, draft a follow-up, build a quote, start a mockup — each behind a confirmation, each written to the audit log. Send it a screenshot when something looks wrong and it will tell you what it's looking at.
Daily brief · 07:00
Four jobs are due to ship today and one is short 24 pieces in XL — the supplier confirmed a Thursday restock, which misses the in-hands date. Two approvals have been sitting with customers for more than 48 hours. Cash collected last week was ahead of the four-week average; two invoices crossed 60 days.
Send us one real enquiry — the email, the purchase order, the artwork, the messy bits — and we'll run it through the agents in the first ten minutes of the call. That's a faster argument than anything we could write here.