What an always-on AI assistant really costs
Everything that has to be paid, built, watched and insured before a monthly price can even be written down. Prepared by Chris Johns' assistant, 26 September 2026. This is a map of burdens, not an offer.
How to read the numbers. measured means a figure taken from Chris' own bills and logs. approx. means a typical market range from memory, to be checked before anyone signs anything. Chris' own statements are marked as his.
1. What "the service" actually is
The thing people see is a phone number that answers like a person, in your language, day and night. Behind it, today, Chris runs: a chat assistant on WhatsApp with a memory of who you are; a second, deeper engine for hard questions; automatic translation between languages; reading and filing his email; phone calls placed and held by an AI voice in Italian; a real browser it can drive on websites that have no API (shopping, marketplaces, logged-in pages); a task board shared between him and the assistants; production boards for film previz that the assistant can read and, with permission, edit; image and video rendering on his own GPU machines and on paid cloud engines; voice notes in and out; a Signal bridge; a room microphone that answers when spoken to. Each of those is a separate machine, subscription, integration and failure point. The price is the sum of them, plus the person who keeps them alive.
2. The cost heads
A. Intelligence: model tokens
- Every message the assistant reads or writes is billed per token, and the bill depends on which model answers. List prices today, per million tokens in/out: top model $10/$50, mid $5/$25, fast $2/$10, small $1/$5. measured
- One user's WhatsApp chat on the top model at high effort cost about $25 a day (roughly $750 a month) before any routing. measured
- At peak, the account was auto-recharging about $45 a day, one receipt every few hours. measured
- After a router that sends ordinary messages to the fast model and only flagged ones to the top model: about $0.09 per ordinary turn, about $0.49 per deep turn. measured
- Memory grows: the assistant carries your history in every call, so the cost per message rises with the months unless someone prunes it.
- Switching models mid-conversation re-writes the cache (about $0.50 each time on the top model). Long documents, images, PDFs and screenshots cost far more than text.
- Prices, limits and model names change without notice; models get retired; a retired model means engineering time to migrate.
- The impatience premium: caps off, deep answers on everything, fast lanes for renders. Chris says he spends over $1,000 a month himself this way. (His statement.)
B. Subscriptions that sit underneath
- Developer seat for the coding assistant that builds and repairs everything: $200 a month per seat. measured
- A second AI provider for personal memory and search (Chris pays $200 a month for that one too). measured
- Voice cloning and speech: roughly $22 to $99 a month by tier. approx.
- Music generation: roughly $10 to $30 a month, credits run out. approx.
- Video engines by the second: around $0.009 to $0.011 per second of video at low resolution, from $0.09 per second and up for HD, 10 to 15 times more; a 30-second spot with re-rolls is real money. measured
- Speech-to-text, image models, translation APIs, mapping, search: each a small line, all of them together not small.
C. Phone, SIM and messaging channels
- iMessage requires an Apple machine (a Mac mini, roughly $600 to $800 once) that stays on, an Apple ID, and a real mobile number on a real SIM from a major carrier. A virtual VoIP number does not register. Carrier plan roughly $30 to $70 a month in the US. approx. If Apple changes something, the whole channel can stop for days.
- WhatsApp, two roads: the official Business API (business verification, message templates, per-conversation fees, no "act as a person") or an unofficial linked device (behaves like a person, breaks Meta's terms, and a ban burns the number and every conversation on it). An operator who takes the second road runs it with strict rate limits, one contact at a time. That risk has a cost: a burned number means a new number, new pairing, lost history.
- VoIP numbers and calls: a number costs about $1 a month, calls about a cent a minute, plus the PBX that holds the call and the AI voice that speaks. approx. Outbound SMS in the US needs campaign registration; in Italy KYC on the number.
- Signal, Telegram, email: each its own bridge, each with its own login that expires and needs a human to re-pair.
- A second client means a second number on every channel. Numbers cannot be shared between clients.
D. Hardware and electricity
- Local rendering and transcription run on GPU workstations: Chris runs two NVIDIA DGX Spark units plus a third machine for the room and local models. A DGX Spark bundle costs about €12,000, plus about €4,000 for the third machine (an EVO-X2). quoted Without them, every render is a cloud bill.
- Electricity: a GPU box at 200 to 400 watts, 24 hours a day, is 150 to 300 kWh a month. At Italian prices that is roughly €45 to €90 a month per machine. approx.
- UPS, disks, backups, a second copy of everything somewhere else, and the day a disk dies.
- Thermal throttling, driver updates, a kernel update that breaks the GPU stack: hours, not euros, but hours cost.
E. Cloud and network
- A small public server as the front door (roughly $6 to $50 a month), domains, TLS certificates, DNS, a private mesh network between machines, log storage, monitoring. approx.
- Uptime is a feature: "reboot-proof" means every service restarts on its own and someone tests that it does.
F. Headless and logged-in browsers
- Sites without an API (marketplaces, shops, portals) are driven through a real browser session that stays logged in. Cookies expire, two-factor codes arrive on a phone, CAPTCHAs appear, layouts change: each of those needs a human at the right moment or the task silently fails.
- Visual browsing (the model looking at screenshots to decide what to click) costs many times a text turn.
- Account bans and rate limits on the client's own accounts are a real liability. Purchases on someone's behalf need a written mandate and a spending cap.
- Residential IPs, proxies, a second browser profile per client: more lines.
G. Integrations: each one is engineering
- Gmail, Calendar, Drive, WhatsApp, iMessage, Signal, phone calls, shared task board, production boards, e-commerce, banking alerts, DistroKid, YouTube: every one took days to build and breaks when the vendor changes an API, a login flow or a policy.
- Rule of thumb from this year: something breaks every week. Maintenance is the largest recurring line after tokens, and it is paid in hours.
H. Permissions, safety and guardrails
- Who may the assistant write to, buy from, delete, publish? Per-contact levels, allowlists, code words, audit logs, a human gate for anything irreversible, spend caps per day, autonomy tiers by task.
- Removing permissions makes it safer and slower (a human has to say yes; that human's time is a cost). Removing guardrails makes it faster and creates overruns: a runaway loop on the top model can spend hundreds of dollars overnight.
- Prompt injection: everyone who messages the assistant's number is a stranger who can try to give it orders. Every inbound message has to be treated as data, never as instruction. That discipline has to be built and tested.
- Secrets: API keys, session cookies, phone pairings, mailbox tokens. Rotation, storage, what happens when a laptop is lost.
- Kill switches and the ability to un-do: backups before every write, versioned files, a way to roll back a bad edit to a client's board.
I. Operations: it never sleeps, so someone has to
- 24/7 uptime, watchdogs, restarts, session resets (the assistant's working memory fills up and has to be rolled over without losing the thread), nightly jobs, a "context clock" so a tired session stops touching client work.
- On-call: when the number goes silent at 3 a.m., who notices and who fixes it.
- Logging and audit: every message sent on the client's behalf recorded, retrievable, deletable on request.
J. Human time: the biggest hidden cost
- Onboarding: days of teaching the assistant the client's people, preferences, language, what it must never do.
- Reviewing: outputs still have to be read by a human before they reach third parties; the assistant is a very fast junior, not a partner.
- Fixing: each vendor change, each new site, each new wish is engineering. Chris' own hours are the real price of his setup.
K. Legal and compliance (Italy / EU): must be checked, not legal advice
- GDPR: the assistant processes third parties' messages, contacts and voice. That needs a lawful basis, a data-processing agreement between client and operator, retention rules, and the fact that the models run on US providers (data transfer terms).
- EU AI Act transparency: people must be told they are talking to an AI; synthetic voice on calls must be disclosed. Obligations phase in from 2026.
- Call recording: consent and disclosure rules differ by country and by who is on the line.
- WhatsApp terms: automation via an unofficial client is against Meta's terms; the client accepts the ban risk in writing or uses the official API with its constraints.
- Acting on someone's behalf: purchases, contracts, messages signed as "the assistant" need a written mandate and limits.
- Confidentiality: a producer's scripts, budgets, cast lists and contracts flow through the system; NDA, access control, and what the model provider may retain.
- Content rights: generated images, music and video, likeness of real actors (previz only), licences of the generation tools.
- VAT and invoicing between an Italian company and a foreign operator.
L. Financial risk
- Auto-recharge means the bill is open-ended by design; the operator carries the overrun until it is noticed.
- Vendor price rises, currency, a model deprecation that forces a costlier one, a credit balance that hits zero and stops every job at once.
- Chargebacks, unpaid months, and the fact that most of the cost is spent before the month is invoiced.
M. Scaling to a second client
- Nothing can be shared: a second client means separate numbers, separate sessions, separate secrets, separate memory, separate audit, a separate browser, and ideally a separate machine. That is a product, not a copy.
- Support: a client who does not live inside the terminal needs a person to answer "why did it do that".
3. Orders of magnitude for one seat
| Line | Monthly, order of magnitude | Confidence |
|---|---|---|
| Model tokens, routed and capped | $100 to $300 | estimated from measured per-turn costs |
| Model tokens, caps off, top model everywhere | $750 and up | measured |
| Developer seat that builds and repairs | $200 | measured |
| Voice, music, video credits | $30 to $200 | usage |
| Phone: real SIM + VoIP + calls | $40 to $100 | approx. |
| Mac mini for iMessage | $600 to $800 once | approx. |
| Cloud front door, domains, monitoring | $20 to $80 | approx. |
| Electricity for one GPU box | €45 to €90 | approx. |
| Human hours: onboarding, review, repair | the largest line; not in dollars yet |
Chris' framing, in his words: a service like this would be in the hundreds of dollars a month, somewhere between $200 and $500 depending on what it has to do, and he spends over $1,000 a month himself because he takes all the caps off. He mentions comparable services (he named Instinct.ai) that are invitation-only for exactly these reasons. Those are his statements, not verified market data.
4. The long list: every line, minute to grand
Written out on purpose so the size of the calculation is visible. Some lines are cents, some are thousands, all of them have to be known, owned and watched. Grouped; 174 items.
Tokens and models
- Input tokens per message
- Output tokens per message
- Cache write tokens (first time a long context is loaded)
- Cache read tokens (every later turn)
- Cache expiry: a pause longer than the cache window re-bills the whole context
- System prompt size (rules, laws, persona) billed on every turn
- Memory files loaded per session, growing monthly
- Tool definitions loaded per turn
- Thinking / reasoning tokens at high effort
- Retries after a timeout or a truncated answer
- Duplicate answers when a channel redelivers a message
- Model routing errors: a trivial question sent to the top model
- Model switch mid-thread rewriting the cache
- Image inputs (a photo costs hundreds to thousands of tokens)
- PDF and document inputs
- Screenshot loops in visual browsing (one screenshot per click)
- Transcription of voice notes (minutes of audio)
- Translation of every inbound and outbound message on multilingual chats
- Summaries written when a session's context is compacted
- Overnight and scheduled jobs that run whether or not anyone asked
- Guard and watchdog jobs that use a model to check the others
- Evaluation runs (testing a change against old conversations)
- Price changes by the provider
- Model deprecations and forced migrations
- Rate limits and 429s: queueing, back-off, lost work
- Provider outages: work redone later
- Auto-recharge threshold and its tax
- A second provider kept as a fallback (and its minimum spend)
- Local models as a cheap lane: hardware, power and the hours to keep them running
Subscriptions and credits
- Coding assistant seat ($200/month)
- Second AI provider with personal memory ($200/month)
- Voice cloning and TTS tier
- Extra voice characters beyond the tier
- Music generation tier and top-up credits
- Video engine credits by the second
- Re-rolls: the 3 to 5 takes behind every kept clip
- Image generation credits
- Speech-to-text minutes
- Search API calls
- Maps and geocoding calls
- OCR for scanned documents
- Email API quota (Google Workspace seat if the volume needs it)
- Calendar and Drive API quotas
- Cloud storage for media (per GB and per download)
- Distribution accounts (music, video platforms)
- Domain registrations (one per brand, one per client)
- Password manager for the operator team
- Monitoring / uptime service
- Error tracking service
- Analytics
Phone and messaging
- Real SIM from a major carrier for iMessage
- Carrier plan monthly fee
- Roaming when the device travels
- A phone or Mac to host that SIM's registration
- Apple ID and two-factor device for it
- VoIP number rental (per number)
- Number porting fees
- Per-minute inbound calls
- Per-minute outbound calls
- International rates (US to Italy and back)
- SMS per message
- US 10DLC campaign registration and vetting
- Italian KYC on numbers
- Emergency-calling compliance on VoIP
- PBX server (the software switchboard)
- Call recording storage
- Voice AI per-minute (speech in, model, speech out)
- Latency engineering so the voice sounds like a person
- Voicemail and missed-call handling
- WhatsApp: official API business verification
- WhatsApp: template approval and per-conversation fees
- WhatsApp: unofficial device pairing and its ban risk
- A burned number: replacement, re-pairing, lost history
- Rate limiting to look human (which slows every task)
- Signal linked device and its re-linking
- Telegram bot or user session
- Email sending reputation (SPF, DKIM, DMARC, warm-up)
- Spam-law compliance for outbound messages
- Contact roster maintenance (who is allowed)
- Per-contact permission levels
- Group chats: reading many people's messages
- Media received: photos, documents, voice notes to store or discard
- Read receipts and typing indicators (small things people notice)
Hardware, power and physical
- GPU workstation(s) purchase
- Mac mini for iMessage
- A machine for local models and the room microphone
- Room microphone and speaker
- Disks and their replacement
- Off-site backup target
- UPS and its battery replacement
- Network switch, cables, a second internet line for failover
- Electricity per machine, 24/7
- Cooling and the room it heats
- Noise (fans in a home or office)
- Depreciation (three years on GPUs, two on phones)
- Warranty and repair turnaround (weeks without a box)
- Physical security of a machine holding everyone's sessions
- Shipping and customs when hardware crosses borders
Cloud, network and security
- Public front-door server
- Bandwidth and egress
- TLS certificates and their renewal
- DNS hosting
- Private mesh network between machines
- Firewalling and exposure reviews
- Secrets storage and rotation (API keys, cookies, tokens)
- SSH keys and who holds them
- Log storage and retention
- Backups: hourly snapshots, restore drills
- Git hosting for the code
- Reboot-proofing every service and proving it
- OS and driver updates (and the ones that break the GPU stack)
- Vulnerability patching
- Incident response when something is exposed
Browsers and web automation
- Persistent logged-in browser profile per client
- Cookie expiry and re-login
- Two-factor codes needing a human
- CAPTCHA appearances
- Layout changes on target sites
- Residential IP or proxy
- Remote screen for a human to step in
- Visual-model cost per screenshot
- Site terms of service for automation
- Client account bans
- Purchases: mandate, spending cap, receipts filed
- Refund and dispute handling
Integrations and engineering
- Each channel bridge (WhatsApp, iMessage, Signal, email, voice, room)
- Each account integration (Gmail, Calendar, Drive, shops, platforms)
- Task board and its API
- Production boards and their guards
- Rendering pipeline and its queue
- Memory system (files, index, pruning)
- Permission system and audit log
- Router between fast and deep models
- Spend caps and alerts
- Context clock and session rollover
- Testing after every change
- Documentation so the next person can fix it
- Vendor API changes (weekly, somewhere)
- Technical debt paid every month
People and process
- Onboarding interviews with the client
- Writing the client's rules and no-gos
- Teaching names, relationships, languages, tone
- Daily review of what went out
- Approval turnaround (the client's own time)
- On-call nights and weekends
- Incident write-ups
- Client support: explaining what happened and why
- Training the client to use the codes and commands
- Quarterly review of permissions and spend
- Offboarding: deleting data, releasing numbers
Legal, compliance and insurance
- Service contract
- Data-processing agreement (GDPR)
- Records of processing and retention schedule
- Data-transfer terms for US model providers
- AI Act transparency notices (chat and synthetic voice)
- Call-recording consent script
- WhatsApp terms risk acceptance
- Written mandate for acting on the client's behalf
- NDA for scripts, budgets and contracts
- Content licences for generated media
- Likeness rights for real people in previz
- Professional liability insurance
- Cyber insurance
- VAT treatment and invoicing across borders
- Tax on hardware and subscriptions
- Legal review time (hours of a lawyer)
Financial
- Overrun buffer (the runaway night)
- Working capital: costs land before the invoice
- Payment processing fees
- Currency swings USD/EUR
- Unpaid or disputed months
- Minimum commitments on vendor plans
- Price increases passed through or absorbed
- The operator's own margin, which has to exist for the service to survive
5. What decides the price: questions to answer first
- Which channels: WhatsApp only, or also iMessage, Signal, email, phone calls, a room microphone?
- May it write to third parties as you, as "your assistant", or only draft for your approval?
- Which languages, and does every message get translated?
- Which accounts does it log into (email, calendar, shops, marketplaces, social) and what may it do there: read, write, buy, publish?
- Spending caps: per day, per task, per month. Who is told when a cap is hit?
- Fast or deep: what share of questions needs the top model, and how patient are you?
- Rendering: images and video at all? On whose hardware? How many per week?
- Uptime: 24/7 with on-call, or business hours with best effort?
- Data: what is kept, for how long, where, and who can read the logs?
- Legal: who is the data controller, who signs the processing agreement, who accepts the WhatsApp-terms risk?
- Exit: what happens to the number, the memory and the history if the service stops?
Until those are answered, any number is a guess, and the risk sits with whoever wrote it down.