Own vs. Buy

Anyone Can Build an AI Agent. Can You Own One?

Building is the easy part, and even that is hard. The real bill arrives after launch, when models change, APIs break, knowledge goes stale and your agent’s answers become your liability. That’s owning it.

Price your ownershipSee day two
Life of an in-house agent
Illustrative
Model retired
Eval regression
API changed
Knowledge stale
Wrong answer
Builder left
Build
Own
Launch
Year 1
Year 2
Year 3
Since you opened this page, owning the build in the calculator below has cost
$0.00
Live, at $521 per day in production. Change the calculator inputs and it recalculates.
Day two never ends

Building Is a Project. Owning Is a Job.

Software research has said it for decades: maintenance, not development, is where the money goes. On average it’s about 60% of a system’s lifecycle cost, and anywhere from 40% to 80% depending on the system. [1]

AI agents add new kinds of upkeep. The model can change under you without a single line of your code changing. [3] And when your agent gets an answer wrong, the answer is yours. [6]

agent-ops.log
Live
Mon 02:14
ERROR
provider Model version retired. Prompts tuned for it now drift.
Mon 09:40
WARN
eval Same test set, new model build: accuracy 84.0% → 51.1% [3]
Tue 11:05
ERROR
crm Upstream API changed a field. Ticket creation failing.
Tue 16:32
WARN
knowledge Sync stalled. Agent answering from an outdated policy.
Wed 08:17
ERROR
guardrail Agent described a fare policy the airline didn’t have [6]
Wed 13:50
WARN
infra Rate limit reached. Conversations queued.
Thu 10:02
INFO
security InfoSec opened a new review of agent data access.
Thu 18:44
WARN
team Engineer who built retrieval resigned. No runbook found.
Fri 03:29
ERROR
cost Token spend over budget again.
Mon 02:14
ERROR
provider Model version retired. Prompts tuned for it now drift.
Mon 09:40
WARN
eval Same test set, new model build: accuracy 84.0% → 51.1% [3]
Tue 11:05
ERROR
crm Upstream API changed a field. Ticket creation failing.
Tue 16:32
WARN
knowledge Sync stalled. Agent answering from an outdated policy.
Wed 08:17
ERROR
guardrail Agent described a fare policy the airline didn’t have [6]
Wed 13:50
WARN
infra Rate limit reached. Conversations queued.
Thu 10:02
INFO
security InfoSec opened a new review of agent data access.
Thu 18:44
WARN
team Engineer who built retrieval resigned. No runbook found.
Fri 03:29
ERROR
cost Token spend over budget again.
Illustrative week. Each line is a category of failure documented in the sources below.
Ownership calculator

Price Your Ownership.

Start with your build plan. Research defaults are marked; your assumptions are yours to change.

4
Your assumption
6
Your assumption
$133,080
Research default: $133,080, the U.S. median for software developers [7]
70%
Research default: 70%, so benefits add the rest [8]
60%
Research default: 60%, within a 40% to 80% range [1]
3
Your assumption
Total cost to ownComputing
$950,571
Build (day one)$380,229
Own (every day after)$570,343
Cost per day in production$521
Engineer-years consumed5.0
The opportunity costThat’s 5.0 engineer-years not spent on your core product. Every sprint on agent infrastructure is a sprint not spent shipping the features that drive growth and revenue.

How it works: loaded cost per engineer is salary divided by the wage share. Build cost is engineers × loaded cost × build time. Ownership cost follows from the maintenance share, so at 60% it is 1.5× the build. People costs only: model usage, hosting, tooling and incident costs would add to the total.

What you actually own

The Agent Is the Small Box.

Google engineers showed that in real-world machine learning systems, the model code is a small fraction of the system. The rest is everything around it, and that’s where the ongoing maintenance cost piles up. [2]

Upkeep
Model upgrades
Providers retire and change models on their schedule.
Quality
Evals and regression tests
Every change needs re-testing against real conversations.
Content
Knowledge sync
Policies change daily. Stale answers are wrong answers.
Systems
Integrations
CRMs, ticketing and order systems ship breaking changes.
Yours
Agent logic
The part you set out to build.
Risk
Guardrails and compliance
What the agent says, you said.
Ops
Monitoring and on-call
Someone gets paged when it breaks at 2 a.m.
Trust
Security reviews
Data access, retention and audit, reviewed again and again.
People
Knowledge in heads
When the builder leaves, so does the context.
Independent research

Don’t Take Our Word for It.

Nine findings from published third-party sources: academic research, market research, a legal ruling, an industry reference and analyst research. Figures are reported as published.

The Cost of Owning an AI Agent

Sources include
Harvard Data
Science Review
O’ReillyBC Civil Resolution Tribunal
Part 1 · Independent studies and public records
Maintenance cost
60%

Maintenance accounts for about 60% of a software system’s lifecycle cost on average, ranging from 40% to 80% by system type. Most of it is enhancement driven by changing requirements.

Source: David Wood, “The 60/60 Rule,” in 97 Things Every Project Manager Should Know, O’Reilly Media. [1]
Model drift
84% → 51%

On the same prime vs. composite task, GPT-4’s accuracy fell from 84.0% in March 2023 to 51.1% in June 2023. The authors conclude that LLM behavior can change substantially in a short time and needs continuous monitoring.

Source: L. Chen, M. Zaharia, J. Zou, “How Is ChatGPT’s Behavior Changing Over Time?” Harvard Data Science Review, 2024. [3]
Build vs. buy success
67%
Bought
Built

AI tools bought from specialized vendors succeeded 67% of the time. Internal builds succeeded one-third as often. The study drew on 150 leadership interviews, a survey of 350 employees and an analysis of 300 public AI deployments.

Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025, as reported by Fortune, Aug. 18, 2025. [4]
Pilot to production
88%

88% of observed AI proofs of concept don’t make it to widescale deployment. For every 33 AI proofs of concept a company launched, only four graduated to production.

Source: IDC research with Lenovo, as reported by CIO.com, Mar. 25, 2025. [9]
Agent liability
C$812.02

A British Columbia tribunal ordered Air Canada to pay damages, interest and fees after its website chatbot gave a customer wrong information about bereavement fares. It rejected the argument that the chatbot was a separate entity responsible for its own statements.

Source: Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, Feb. 14, 2024. [6]
Part 2 · Analyst perspective: Gartner
Project cancellations
40%+

Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.

Source: Gartner, press release, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 25, 2025. [5]
Time to production
8.2 months

Gartner found that “on average, only 48% of AI prototypes make it into production,” and that going from AI prototype to production takes “on average 8.2 months.”

Source: Gartner, Survey Shows How Generative AI Puts Organizational AI Maturity to the Test, Leinar Ramos, Anthony Mullen, Bern Elliot, 23 April 2024.
Initial investment
Build
Buy
$0
$1.5M
$3M
$4.5M

Gartner puts the initial investment to build AI agents at $2.5M to $4.3M, against $1.5M to $2.7M to buy.

Chart created by Maven AGI based on Gartner research. Source: Gartner, How to Calculate the Value and Cost of AI Agents, [Analyst names], [Publication date].
The ongoing burden
Model upkeep
Rare AI talent
Team capacity
Compliance audits

Building means owning the model upkeep a vendor would handle, hiring rare AI talent and carrying heavier compliance audits. Gartner warns against overestimating your team’s capacity to sustain it.

Sources: Gartner, How to Calculate the Value and Cost of AI Agents; Choosing Between an Agent Harness or Framework to Build AI Agents; 2027 Planning Guide for Software Architecture and Integration. [Analyst names], [Publication dates].

Methodology. Figures are reported as published by each source and have not been adjusted by Maven AGI. Exhibits are illustrations created by Maven AGI from the cited figures. Full citations appear under Sources.

Gartner research publications consist of the opinions of Gartner’s research organization and should not be construed as statements of fact. GARTNER is a registered trademark and service mark of Gartner, Inc. and/or its affiliates in the U.S. and internationally and is used herein with permission. All rights reserved.

Who carries what

Own the Experience. Not the Plumbing.

Ongoing work
If you own it
If you buy Maven AGI
Model upgrades and deprecations
Your engineers
Maven AGI
Regression testing after every change
Your engineers
Maven AGI
Knowledge freshness
Your engineers
Maven AGI
Integration upkeep as APIs change
Your engineers
Maven AGI
Guardrails, compliance and audit
Your engineers
Maven AGI
Security reviews and data controls
Your engineers
Maven AGI
Monitoring and on-call
Your engineers
Maven AGI
Your policies, brand voice and customer experience
You
Still you
Buying isn’t a black box
Build what’s yours on top.
Through the Maven developer portal, customers have built their own chat experiences and custom integrations such as WhatsApp. You get the control of a build, on a platform someone else keeps running.

Keep Your Engineers on What Makes You Different.

Maven AGI gives you agents you control, without the job of keeping them alive. Our customers go live in days or weeks, not months.

Book a demoRun the numbers again
Sources
1. David Wood, “The 60/60 Rule,” in 97 Things Every Project Manager Should Know, O’Reilly. oreilly.com
2. D. Sculley et al. (Google), “Hidden Technical Debt in Machine Learning Systems,” NeurIPS 2015. neurips.cc
3. L. Chen, M. Zaharia, J. Zou, “How Is ChatGPT’s Behavior Changing Over Time?” Harvard Data Science Review, 2024. hdsr.mitpress.mit.edu
4. MIT NANDA, “The GenAI Divide: State of AI in Business 2025,” as reported by Fortune, Aug. 18, 2025. fortune.com
5. Gartner, press release, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 25, 2025. Coverage: sdtimes.com
6. Moffatt v. Air Canada, 2024 BCCRT 149: C$812.02 in damages, interest and fees. mondaq.com
7. U.S. Bureau of Labor Statistics, Occupational Outlook Handbook: software developers’ median annual wage, May 2024. bls.gov
8. U.S. Bureau of Labor Statistics, Employer Costs for Employee Compensation, June 2026: wages and salaries were 70.0% of private-industry employer costs. bls.gov
9. IDC research with Lenovo: for every 33 AI proofs of concept, four reached production, as reported by CIO.com, Mar. 25, 2025. cio.com