How to Build an AI Agent, and When to White Label One Instead

What an AI agent is made of

OpenAI's practical guide puts it in three parts: a model that does the reasoning, tools the agent can call to fetch data or take actions, and instructions that set its behaviour and guardrails. It suggests building an agent only for workflows that need judgement, have rules too tangled to maintain, or rely on unstructured language, and says a deterministic solution may suffice otherwise. Its general recommendation is to maximise a single agent's capabilities first, since a single agent with tools is often sufficient ¹.

Anthropic draws a line between workflows, where code decides the steps, and agents, where the model decides. It advises finding the simplest solution possible and warns that agentic systems often trade latency and cost for better task performance, so the trade should be weighed case by case ².

For a customer-facing agent, the three parts turn into a longer list once it has to reach real people: a voice agent needs telephony, speech to text and text to speech, a WhatsApp or Instagram agent needs Meta's platforms and their approvals, and any agent for clients needs a knowledge base, a way to hand off to a person, logs, a login for the client and a bill at the end of the month.

What a prototype takes

With a current coding assistant a single developer can wire a model to a knowledge base and a chat widget in a day or two, and the published evidence on AI-assisted coding supports that for defined tasks: 55.8% faster in a randomised Copilot experiment and 40% faster on writing tasks with ChatGPT ^15,16. The tools cost little. Claude's individual Pro plan is $17 a month on annual billing, or $20 monthly, and the Max plan starts at $100 ³.

The same evidence has a second half, because Veracode found 45% of AI-generated code samples carried a security vulnerability, and a scan of 1,645 apps built with one vibe coding tool found 170 exposing user data ^18,19. A client agent holds call recordings, phone numbers and conversations, which is the data those flaws expose. Our vibe coding page covers that evidence in full.

What it costs to run: list prices

These are list prices from the vendors' own pages, read on 14 October 2026. For the model, Anthropic lists Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, Sonnet 5.5 at $2 and $10, and Opus 5.5 at $4 and $20. OpenAI lists gpt-5.4-nano at $0.20 and $1.25, gpt-5.4-mini at $0.75 and $4.50, and gpt-5.5 at $5 and $30 ^3,6.

For voice, Deepgram lists streaming speech to text at $0.0048 a minute and its Aura-2 text to speech at $0.030 per 1,000 characters. Twilio lists US inbound local calls at $0.0085 a minute, outbound at $0.014, a local number at $1.15 a month and call recording at $0.0025 a minute. ElevenLabs' agent product is reported at $0.08 a minute on every plan, with model and telephony extra; its own page did not load for us ^4,5,7.

For messaging, Twilio charges $0.005 per WhatsApp message on top of Meta's fees, and Meta's rules since July 2025 make replies inside the 24-hour service window free while marketing templates are charged ⁸. Hosting a small web app on Vercel's Pro plan is $20 a month per developer seat with $20 of usage credit included ²⁰.

What it costs to run: people

The US Bureau of Labor Statistics puts the median pay of software developers at $134,040 a year, or $64.44 an hour, in May 2025, with 1,905,400 jobs and 10% growth projected to 2035 ⁹. At that wage, Convocore's $199 a month White Label plan equals about 3.1 hours of one developer's time, by our division, before employer costs ^9,11.

Building is the smaller part of what a developer will spend on a product, since reviews of software economics going back to the 1980s put maintenance at 60% to 80% of total lifecycle cost, with some estimates above 90%, and although the figures are old and the studies uneven, no study we found puts it below half ¹⁰.

What changed under working agents in the last twelve months

An agent that worked in October 2025 has had to change several times to still work today, because OpenAI's deprecation calendar lists twelve shutdown dates in 2026, and on 23 October alone it retires gpt-4, gpt-4-turbo, gpt-3.5-turbo, gpt-4.1-nano, the gpt-4o-2024-05-13 snapshot, o1, o3-mini, o4-mini and gpt-image-1. It shut the Assistants API on 26 August, the realtime preview models on 7 May, and will retire the first GPT-5 snapshots on 11 December 2026 ¹².

Meta ended the WhatsApp On-Premises API on 23 October 2025, after which numbers still registered on it could not send or receive messages; new numbers had been Cloud API only since July 2024 ¹³. An app that messages Instagram accounts it does not own needs Advanced Access, which requires Meta's App Review, and Meta has required Business Verification for Advanced Access since 2023; reports of verification taking from 30 days to over three months are common but there is no official figure ¹⁴.

US text messaging needs A2P 10DLC registration through The Campaign Registry. Twilio's support pages, as summarised in our search, list a $46 one-time fee for a standard brand including secondary vetting, a $15 campaign vetting fee, and monthly campaign fees of $1.50 to $10 depending on use case, per brand and campaign. We could not open the pages themselves ²¹. And carriers fail: third-party trackers of Twilio's status page logged a 19 hour SMS incident in Italy in March 2026, a 159 hour one across several European carriers in May, and a 57 hour voice incident in Brazil in June ²².

Build or white label: the evidence

The one study that compares the two routes is about enterprises. MIT's NANDA initiative, drawing on about 150 interviews, a survey of 350 employees and 300 public deployments, reported that AI tools bought from specialized vendors or built with partners succeeded about 67% of the time and internal builds roughly a third as often, as reported by Fortune ¹⁷. It is not about agencies and reaches us through a press report, so treat it as a direction.

The arithmetic above gives the rest: compute is close to a wash, the platform fee is about three hours of developer time a month, and most of the work a built agent needs comes after the build, much of it forced by provider calendars the agency does not control. What an agency gives up by white labelling is ownership of the code, and what it takes on is dependence on the platform's own prices and survival.

If you build: a checklist from the evidence

None of this says do not build, it says what a build has to include before a client's customers rely on it. The checklist below is drawn from the sources on this page and from the model makers' own guides ^1,2.

• Start with one agent and tools; add agents only when it fails to follow instructions or picks the wrong tools ¹.
• Pin model versions and read the deprecation calendar every month ¹².
• Register for A2P 10DLC before promising a client SMS, and start Meta Business Verification before promising Instagram ^14,21.
• Keep development and production separate, with backups you have restored at least once ¹⁹.
• Have a person review database access rules, authentication and anything that handles payments ¹⁸.
• Log every conversation and give the client a way to read it and to reach a human.
• Budget maintenance at more than the build ¹⁰.

What the evidence does not show

• No study measures what agencies spend to build and run their own agent platforms, or how many succeed. The MIT comparison is about enterprises and reaches us through Fortune ¹⁷.
• Prices are list prices on the day we read them and change often. Discounts, volume tiers and regional prices differ ^3,4,5,6,8,20.
• ElevenLabs' agent page and Twilio's A2P fee pages did not load for us, so those figures come from search summaries ^7,21. The outage durations come from third-party trackers ²².
• The maintenance share figures are old and the studies are uneven ¹⁰. The developer wage is a US median and ignores employer costs ⁹.
• The component counts for a client-ready agent and the per-minute worked example are our own, not from a source.
• Convocore's figures are from its own catalog. Its agents can be exported as templates, but we have not tested moving one to another platform ¹¹.

What we would do

If you are testing whether clients will pay for an AI agent, build the prototype yourself with a coding assistant, because the evidence says that is fast and the running cost is cents ^15,16, and before the first client's customers talk to it, have someone who can read the code review how it stores data and who can reach it ^18,19.

If you want to sell agents to several clients and you do not employ engineers, the arithmetic favours white labelling: the platform fee is a few hours of developer time, the integrations and migrations are the platform's problem, and you keep the client, so Convocore's White Label plan is $199 a month for up to 20 client sub-accounts with $60 of usage included, voice at $0.02 a minute plus providers at cost, and agents you can export as templates ¹¹, while if you employ engineers and want to own a product, build it and budget for the calendar above.

Method and limits

• Every number on this page comes from a source we opened and read on 14 October 2026, except where a source note below says we could only read a search summary.
• Sources are labelled primary (read at the original publisher), commissioned (paid for by a party with a commercial interest in the result), secondary (a report of the original) or forecast. Vendor price pages are labelled primary because the vendor is the authority on its own price.
• Where sources disagreed or a figure could not be traced to an original, we left it out. Arithmetic of our own is labelled as ours wherever it appears.
• Convocore sells the white label platform discussed here and has a direct commercial interest in this comparison. We have set out the case for building as well as for buying. If you find an error, write to support@convocore.ai and we will correct it.

Sources

1. A practical guide to building agentsPrimary source
   OpenAI. 2025.
   We read the PDF. OpenAI sells the models and the agents SDK it describes.
2. Building effective agentsPrimary source
   Anthropic. 19 December 2024.
   Anthropic sells the models discussed. The page notes that much of the tooling has changed since publication.
3. Claude pricing (API and plans)Primary source
   Anthropic. Read 14 October 2026.
   List prices in US dollars. Haiku 5.5 prices rise for prompts over 100,000 tokens.
4. Deepgram pricingPrimary source
   Deepgram. Read 14 October 2026.
   Pay as you go list prices. Nova-3 streaming speech to text and Aura-2 text to speech.
5. Twilio Voice pricing, United StatesPrimary source
   Twilio. Read 14 October 2026.
   Pay as you go list prices for local and toll-free calls, numbers and recording.
6. OpenAI API pricingPrimary source
   OpenAI. Read 14 October 2026.
   Standard tier list prices. Cached, batch and long context rates differ.
7. ElevenLabs agents pricing, as reportedSecondary report
   Various comparison sites. 2026.
   ElevenLabs' own agent pricing page returned an error. Several third-party pages agree on $0.08 a minute with model and telephony extra; some are published by competitors.
8. WhatsApp pricing (Twilio)Primary source
   Twilio. Read 14 October 2026.
   Twilio's $0.005 per message fee. Meta's own fees by category and country are in Meta's rate cards, which this page summarises in a calculator.
9. Software Developers, Quality Assurance Analysts, and Testers, Occupational Outlook HandbookPrimary source
   US Bureau of Labor Statistics. May 2025 pay data, 2025 to 2035 projections; read 14 October 2026.
   Median wage across the combined occupation. Employer costs are not included.
10. Measurements for managing software maintenance (review of maintenance cost studies)Secondary report
    Academic review, via AMiner. Various, 1979 to 2000.
    Read through a search summary. The 60% to 80% range is the most consistently reported across the studies the review cites; individual estimates run from about 50% to over 90%, and most are decades old.
11. Convocore pricing: White Label plan and voice costsPrimary source
    Convocore. Read 14 October 2026.
    $199 a month, 20 client sub-accounts, 60,000 credits ($60 of usage) included, 10 concurrent call lines, 3 phone number slots; voice at a $0.02 a minute platform fee plus the listed provider rates. Agents can be exported as templates. This is our own product.
12. DeprecationsPrimary source
    OpenAI developer documentation. Read 14 October 2026.
    Dates and model names as listed on the page on the day we read it. The page changes.
13. On-Premises API sunsetPrimary source
    Meta for Developers. Read 14 October 2026.
    States the final version expired on 23 October 2025 and that the API can no longer send messages.
14. App Review for Instagram APIPrimary source
    Meta for Developers. Read 14 October 2026.
    Marks App Review as required for tech providers serving multiple businesses. The page gives no review time. The Business Verification requirement and the reported durations come from a 2023 Meta developer post and forum reports read through a search summary.
15. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (Peng, Kalliamvakou, Cihon, Demirer)Primary source
    arXiv 2302.06590. 13 February 2023.
    95 professional developers, randomised, one task. Three of the four authors worked at GitHub or Microsoft.
16. Experimental evidence on the productivity effects of generative artificial intelligence (Noy, Zhang)Primary source
    Science. July 2023.
    453 college-educated professionals, randomised, writing tasks.
17. MIT report: 95% of generative AI pilots at companies are failing (on The GenAI Divide: State of AI in Business 2025)Secondary report
    Fortune, reporting MIT NANDA. 18 August 2025.
    We read Fortune's article, not the MIT report. Enterprise deployments. No exact figure for internal builds.
18. AI-Generated Code Poses Major Security Risks in Nearly Half of All Development Tasks (2025 GenAI Code Security Report)Commissioned research
    Veracode. 30 July 2025.
    Vendor research by a company that sells application security software. 80 tasks chosen for weakness potential, more than 100 models.
19. Lovable-built apps found exposing user data (scan of 1,645 applications), and the Replit agent incidentSecondary report
    Slashdot and TechTarget. 30 May 2025 and 22 July 2025.
    News reports. The Lovable scan was done by employees of a competitor. The Replit incident (https://www.techtarget.com/it-infrastructure/news/366627829/Replit-AI-agent-snafu-shot-across-the-bow-for-vibe-coding) was a single case on a side project.
20. Vercel pricingPrimary source
    Vercel. Read 14 October 2026.
    Pro plan at $20 a month per developer seat with $20 of included usage. Usage beyond that is billed.
21. A2P 10DLC fees (Twilio support articles)Secondary report
    Twilio, via search summary. Undated.
    Both Twilio support pages returned errors, so the fees come from a search summary of them: $46 one-time for a standard brand including secondary vetting, $15 per campaign vetting, $1.50 to $10 a month per campaign. Carrier pass-through fees are separate.
22. Twilio outage historySecondary report
    IsDown and UptimeRobot, aggregating Twilio's status page. 2026.
    Third-party trackers read through a search summary. Twilio's own status page is the authoritative record and we did not read its full history.

Questions

Related