Triage and urgency
How urgent each item is, from routine to act now.
- Start with
- Myles-150M
- Step up to
- Myles-400M, for long messages and documents
- Runs on
- Laptop
- Why pick it
- Puts the risky items first, on hardware you already have.
Use cases
Every decision is a question with options. The model scores each option, keeps the ones it is sure of and hands the rest to a person.
How urgent each item is, from routine to act now.
Which team or queue should take it, and when it could belong to two.
Whether a message, payment note or listing looks like a scam.
Whether text contains personal data, and which kind.
Whether a claim, application or request meets your written rules.
What the person is asking for, from your list of request types.
Whether a message is a complaint, so it is handled as one.
Whether a message suggests the person may need extra care.
Whether each claim in a draft is supported by its source.
Whether it is safe for an agent to take the next action.
Whether a post or comment breaks your rules, and which rule.
Which kind of document this is, from your list.
Whether two records describe the same person, claim or item.
Whether a fraud or financial crime alert should be closed, kept or escalated.
Whether a photo or scan passes a check: a receipt, a product, damage, a reading.
Flag smishing that copies your brand, so you can warn customers and report the sender.
Spot cancellation intent in chats and emails, so retention teams reply in time.
Send broadband, mobile and billing faults to the right engineers without a manual sort.
Catch complaints that never use the word, so they are logged and answered on time.
Notice signs of illness, bereavement or money worries so the right team gets involved.
Check notes before they are saved, so payment details never sit in the customer record.
Put reports that affect many customers or vulnerable users at the top of the queue.
Check a customer's photo before booking an engineer, and send the unsure ones to a person.
Pause payments that look like impersonation or investment scams, and ask the customer to check.
Clear the obvious false alarms so investigators spend their time on the real cases.
Find complaints in everyday messages, so each one reaches the right team on time.
Spot early signs of money trouble, so support is offered before arrears build.
Check each application against your written criteria and pass the borderline ones to an underwriter.
Find repeat applications that differ by a spelling or an address, before both are approved.
File payslips, statements and proofs of address in the right place the moment they arrive.
Check transcripts before they are stored or shared, and hold back the ones with card details.
Compare the claim with the wording and send anything unclear to a claims handler.
Check that the photo shows the damage described, and route the doubtful ones for review.
Put escape of water, fire and vulnerable customers first, before the routine claims.
Catch the same loss claimed twice, even when the details are written differently.
Send motor, home and travel claims to the right team on the first touch.
Notice when a policyholder may need more time or support, and alert the right handler.
Read the receipt and check the total against the claim before it is paid.
Sort correspondence to housing, council tax, planning and the other services that answer it.
Check each application against the published rules and pass the borderline ones to an officer.
Bring messages that mention risk to a child or an adult to a safeguarding lead first.
Recognise requests for information in general correspondence, so the statutory clock starts on time.
Check reports and responses before publication, and hold back anything that names a person.
Recognise complaints among general enquiries, so each follows your complaints procedure.
Sort evidence uploads into the right case file without anyone opening each one.
Order the referral inbox so the ones the referrer marked urgent are read first.
Send each referral to the clinic that runs the service, rather than a shared inbox.
Check messages and exports for names, patient numbers and dates of birth before they leave.
Check each referral against the service's acceptance criteria and return incomplete ones promptly.
Sort appointment messages so the booking team can act on each without reading it twice.
Find duplicate records created by spelling or address differences, for a person to merge.
Flag messages that suggest someone may need extra help, so staff can respond with care.
Check the reason, dates and condition against your returns policy, and pass exceptions to a person.
Check returned or listed items from a photo, and send the unsure ones to a person.
Start the right journey from the first message, with no menu for the customer to navigate.
Hold back reviews that break your rules before they are published.
Hold orders that look like card testing or account takeover for a person to check.
Match a photo to the right product from your catalogue, with a probability for each.
Pick out complaints from order queries so they are answered under your complaints process.
Spot customers who may need priority services, from what they write in their messages.
Recognise complaints among billing queries, so each is logged and answered within the deadline.
Put reports of gas smells, exposed wiring or flooding in front of a person immediately.
Send meter, tariff and payment questions to the team that can resolve them first time.
Check a submitted meter photo before the reading is used, and ask again when unsure.
Recognise what the customer needs from their first message and start the right process.
Check proof of delivery photos, and flag the ones that show no parcel or the wrong door.
Send each customer query to the depot or team that holds the parcel.
Check a damage claim against your terms of carriage and pass unclear ones to a person.
Put reports that stop vehicles or affect many deliveries at the top of the list.
Compare a label photo with the order record before the parcel leaves the depot.
Recognise the problem from the first message and start the right claim or trace.
Check each part from a line camera photo, and send the unsure ones to an inspector.
Send fault reports to the engineer or line that owns the equipment.
Bring notes that describe a safety risk to a supervisor before the next shift starts.
Check certificates against your specification and hold back materials that do not clearly comply.
Link new defect reports to the open ticket they repeat, so effort is not duplicated.
Sort incoming documents into the right system without anyone opening each one.
Compare a clause with your playbook position and send departures to a lawyer.
Send new enquiries to the right practice area on arrival, rather than a shared inbox.
Check documents before disclosure or sharing, and flag the ones that name individuals.
Recognise whether an enquiry opens a new matter, so intake and conflict checks start correctly.
Check a summary against the documents it cites before it reaches a client.
Find the same organisation under different names and addresses, for conflict checks and records.
Apply your own community rules at volume, and send the close calls to moderators.
Flag accounts and listings that look like scams before users lose money to them.
Hold back harassing comments for review before they reach the person they target.
Check adverts against your policy before they run, and pass borderline ones to a reviewer.
Check uploaded images against your rules and send the unsure ones to a moderator.
Put reports that describe a threat of harm in front of a person immediately.
Check documents, emails and pages before an agent reads them and acts on what they say.
Check each claim in a generated answer against the passage it came from.
Score the risk of each action an agent is about to take, and ask a person first.
Check prompts and tool results for personal data before they reach a model or a log.
Send each request to the right tool or workflow, quickly and cheaply.
Check a reply against your product's rules before the user sees it.
Check each request against what your agent is allowed to do, and hold back the rest.
Myles-22M
A browser or a phone
27 ms median decision on one desktop CPU thread
Myles-G Edge By request
Laptop- and phone-class GPUs
21 ms median decision on one GPU
Myles-150M
A laptop
196 ms median decision on one desktop CPU thread
Myles-400M
The cheapest server you have
About 500 ms median decision on one desktop CPU thread
Myles-G 12B
One workstation GPU
31 ms median decision on one GPU
Myles-G 31B
One data-centre GPU
48 ms median decision on one GPU
Myles-G 12B
About £1.05 per million typical decisions
31 ms median decision on one GPU
Myles-G 31B
About £3.50 per million typical decisions
48 ms median decision on one GPU