Slide 1
Slide 1 text
The AI-savvy operating model
towards a humane, effective, financially-transparent way of
working for AI-enhanced knowledge work
Matthew Skelton, Conflux - co-author of Team Topologies
DevOpsDays Singapore 2025 | 2025-05-14
K38
Photo by Barbara Zandoval on Unsplash
Slide 2
Slide 2 text
2
Matthew Skelton
holistic innovation
Originator of Adapt Together™ by Conflux
Co-author of Team Topologies
matthewskelton.com
CEO/CTO at Conflux - confluxhq.com
Slide 3
Slide 3 text
How can we empower teams of humans
+ AI agents to deliver quickly, safely, and
compliantly with high-fidelity domain
knowledge and visibility of the data
sources & results, plus the agency to
address problems caused?
3
Slide 4
Slide 4 text
(we already know)
4
Slide 5
Slide 5 text
5
● Empowered teams
● No hand-offs
● Ongoing stewardship
● Clear boundaries
● Defined guardrails and specifications
● Active knowledge diffusion
Slide 6
Slide 6 text
The near-future AI advantage (?)
My perspective on AI (and work)
How to trust and organize AI
The AI-savvy operating model
6
Slide 7
Slide 7 text
The near-future AI advantage (?)
7
Slide 8
Slide 8 text
8
“AI-savvy” operating model
Slide 9
Slide 9 text
9
Traditional AI: pattern matching, huge
data volumes, temporal correlation
Generative AI: next-token guess,
content generation, option generation
Slide 10
Slide 10 text
10
“Learning” but no
understanding with GenAI
Slide 11
Slide 11 text
11
Savvy: don’t pretend or
forget that GenAI doesn’t
understand ⚠
Slide 12
Slide 12 text
12
Savvy use cases for AI:
● Pattern extraction from huge data
● Speed up content generation (incl.
code) and workflows
● Augment human decision-making
● Agent-based processes
● …
Slide 13
Slide 13 text
13
Do like AI and look
backwards!
Slide 14
Slide 14 text
14
Cloud birthed SaaS, leaving
older models behind
(likewise GenAI)
Slide 15
Slide 15 text
15
Agentic AI - devise 230 new product
variations and:
● Generate code, deploy
● Test with synthetic users
● Find the best combination
● Data-driven product fit in hours
A/B testing on steroids!
Slide 16
Slide 16 text
16
but…
Slide 17
Slide 17 text
17
LLMs approaching limits?
Slide 18
Slide 18 text
18
Serious security problems
Slide 19
Slide 19 text
19
Data
Science
Data
Eng
ML Ops
Slide 20
Slide 20 text
20
https://web.devopstopologies.com/
2013
Slide 21
Slide 21 text
End-to-end responsibility for
service outcomes is essential
to avoid harm and deliver
effectively:
use empowered teams
21
Slide 22
Slide 22 text
My perspective on AI
(and work)
22
Slide 23
Slide 23 text
23
1998
My first “AI” system:
backpropagation neural network
for weather data (no, it didn’t work properly!)
BSc, University of Reading
Slide 24
Slide 24 text
24
https://www.linkedin.com/in/promarkbishop/
Slide 25
Slide 25 text
25
2000
MSc in brain science at the
University of Oxford
Research into dyslexia and
Alzheimer’s disease
Slide 26
Slide 26 text
26
2001
Software for MRI brain imaging
machines
Slide 27
Slide 27 text
27
2011
Organizational architecture
and new techniques & tools
for adopting cloud and
Continuous Delivery at
Trainline (UK)
Slide 28
Slide 28 text
28
https://blog.matthewskelton.net/2013/10/22/what-team-s
tructure-is-right-for-devops-to-flourish/
2013
DevOps Topologies
Slide 29
Slide 29 text
29
https://web.devopstopologies.com/
2013
Slide 30
Slide 30 text
Team Topologies
Organizing business and
technology teams for fast flow
Matthew Skelton & Manuel Pais
IT Revolution Press, September 2019
Order via stores worldwide:
teamtopologies.com/book
200k+ copies sold to date in 5 languages
30
Slide 31
Slide 31 text
31
Slide 32
Slide 32 text
Team Topologies paraphrased:
“Given the need for
ongoing stewardship of
long-lived services, how
can we realistically arrange
the flow of value to be
rapid, safe, sustainable?”
32
Slide 33
Slide 33 text
Start with the need for ongoing
evolution of long-lived digital
services meeting user needs…
… and work backwards from there
(don’t start with the technology)
33
Slide 34
Slide 34 text
How to trust and organize AI
34
Slide 35
Slide 35 text
35
Assume we have a collection
of AI agents responding to
some kind of input from
humans…
Slide 36
Slide 36 text
36
Savvy: don’t pretend or
forget that GenAI doesn’t
understand ⚠
Slide 37
Slide 37 text
37
Humans set the context,
goals, guardrails, and
execution constraints in a
repeatable and traceable way
(this is programming)
Slide 38
Slide 38 text
38
“JSON Structure is
a data structure
definition language
that enforces strict
typing, modularity,
and determinism.”
json-structure.org
{
"$schema": "https://json-structure.org/meta/extended/v0/#"
,
"$id": "https://example.com/schemas/product"
,
"$uses": ["JSONStructureAlternateNames"
, "JSONStructureUnits"
],
"type": "object",
"name": "Product",
"properties": {
"id": {
"type": "uuid",
"description": "Unique identifier for the product"
},
"name": {
"type": "string",
"maxLength": 100,
"altnames": {
"json": "product_name"
,
"lang:en": "Product Name"
,
"lang:de": "Produktname"
}
},
"price": {
"type": "decimal",
"precision": 10,
"scale": 2,
"currency": "USD"
},
...
"required": ["id", "name", "price", "created" ]
}
Slide 39
Slide 39 text
39
Cursor / Cline rules
Constrain and guide LLM-based code generation
...
3. Consistency Across Codebases
- Maintain uniform coding conventions and naming schemes across all languages used within a project.
Project Context & Understanding
1. Documentation First
- Review essential documentation before implementation:
- Product Requirements Documents (PRDs)
- README.md
- docs/architecture.md
- docs/technical.md
- tasks/tasks.md
- Request clarification immediately if documentation is incomplete or ambiguous.
2. Architecture Adherence
...
https://gist.github.com/ruvnet/7d4e1d5c9233ab0a1d2a66bf5ec3e58f
Slide 40
Slide 40 text
40
How would we be able to
trust a set of AI agents?
Slide 41
Slide 41 text
41
How are we be able to trust a
set of people?
Slide 42
Slide 42 text
42
RRR:
regular, repeatable results
Slide 43
Slide 43 text
43
plus:
shared language, dashboards,
effective boundaries, data
provenance, decision heuristics,
safe-to-optimize metrics, nimble
governance, psychological safety
Slide 44
Slide 44 text
44
● Guardrails
● Ongoing domain context
● Good boundaries
● Ongoing stewardship and
responsibility
Slide 45
Slide 45 text
45
Teams of humans and AI
agents with context,
guardrails, stewardship,
boundaries, etc…
Slide 46
Slide 46 text
46
Good boundaries for:
● Context
● Security
● Resilience
Slide 47
Slide 47 text
Define guardrails and boundaries
for code gen and service specs:
domains, security, nuances of
terminology, assumptions,
algorithms, biases, etc.
(Hint: this has always been needed!)
47
Slide 48
Slide 48 text
The AI-savvy operating model
48
Slide 49
Slide 49 text
49
Cloud removed IT
infrastructure as a blocker
Slide 50
Slide 50 text
50
AI removes typing,
generating options, and
prototyping as blockers
Slide 51
Slide 51 text
51
Humans understand the
organizational intent and
specify guardrails & goals
Slide 52
Slide 52 text
52
How can we define and align
responsibility for service
outcomes when the code for a
service is generated by AI?
Slide 53
Slide 53 text
Team Topologies paraphrased:
“Given the need for
ongoing stewardship of
long-lived services, how
can we realistically arrange
the flow of value to be
rapid, safe, sustainable?”
54
Slide 54
Slide 54 text
55
Multiple, independent flows, fractally
Slide 55
Slide 55 text
Respect Conway’s Law
(aka ‘sociotechnical
mirroring’)
56
Slide 56
Slide 56 text
Clear ongoing
stewardship of
services and systems
57
Slide 57
Slide 57 text
Stream-aligned teams
have end-to-end
responsibility for a
service (You Build It,
You Run It)
58
Slide 58
Slide 58 text
Platforms improve flow
and reduce extraneous
cognitive load
59
Slide 59
Slide 59 text
Teams are small (~9),
slowly changing, with
‘aligned autonomy’
60
Slide 60
Slide 60 text
Teams are empowered
to sense and adjust
boundaries to improve
flow on a frequent basis
61
Slide 61
Slide 61 text
62
With knowledge work, we’re
fundamentally concerned with
the fidelity of representation
of intent (in code, writing, etc.),
so ongoing domain
knowledge, clear boundaries,
and stewardship are essential.
Slide 62
Slide 62 text
63
Architecture for fast flow
resembles an ecosystem of
loosely-coupled
independently-viable services
with clear boundaries and
ownership aligned to the flow
of business value.
Slide 63
Slide 63 text
64
3EO:
Entrepreneurial
Ecosystem
Enabling
Organizing
https://www.boundaryless.io/3eo-framework/
Slide 64
Slide 64 text
How might AI tools help
leadership decision-making?
65
Slide 65
Slide 65 text
TeamOS
66
teamos.is
Disclosure: Matthew Skelton has invested
personally in the company behind TeamOS
Slide 66
Slide 66 text
67
TeamForm
teamform.co
Slide 67
Slide 67 text
68
CodeScene
codescene.com
Slide 68
Slide 68 text
69
“The work is delivered in many small
changes that are uncoordinated to enable
flow. … Management’s job is to provide
context, prioritization and to coordinate
across teams. Lending resources if needed
across teams to unblock things. … It works
well within a high trust culture.”
Adrian Cockcroft https://mastodon.social/@adrianco/111174832280576410
Technology strategy advisor, Partner at OrionX.net (ex Amazon Sustainability, AWS, Battery Ventures, Netflix, eBay, Sun Microsystems, CCL)
Slide 69
Slide 69 text
If we have clear boundaries for
flow, with limited interactions,
how do we create alignment?
How do we learn from each
other at pace?
70
Slide 70
Slide 70 text
5 DevOps principles: CALMS
● Culture
● Automation / AI
● Lean
● Measurement
● Sharing
71
Slide 71
Slide 71 text
72
Multiple, independent flows, fractally
Slide 72
Slide 72 text
73
Active diffusion of knowledge
across team boundaries
Slide 73
Slide 73 text
75
Future:
AI-powered tools to detect
duplication, uncertainty,
waste, innovation, etc. and
diffuse learning
Slide 74
Slide 74 text
76
internal conferences
guilds
Communities of Practice
lunch & learn
public blogs
Slide 75
Slide 75 text
77
https://internaltechconf.com/
Internal Tech
Conferences
Victoria Morgan-Smith
and Matthew Skelton
Slide 76
Slide 76 text
78
“This initiative around internal
conferences has been the single most
effective thing to align business and
technology that I have seen in this
organization”
– Murray Hennessey, CEO, (UK retail co)
Slide 77
Slide 77 text
79
https://internaltechconf.com/
Internal Tech
Conferences
Victoria Morgan-Smith
and Matthew Skelton
It’s o p te !
Slide 78
Slide 78 text
80
“The way that the Conflux crew used
their active knowledge diffusion
approach to seek out and champion
good practices was a real revelation to
us at TELUS and helped to shift thinking
around how we innovate and share
successes.”
– Steven Tannock, Director, Architecture
(Platform Technology & Tools) at TELUS Digital
Slide 79
Slide 79 text
81
Slide 80
Slide 80 text
82
Thriving organizations, delivering at speed™
Create alignment, trust, and engagement across your
organization whilst delivering at pace with fast flow.
adapttogether.info
Slide 81
Slide 81 text
Combine ‘architecture for flow’
(via Team Topologies) with explicit
‘active knowledge diffusion’
across flow boundaries to create
trust, alignment, and learning.
83
Slide 82
Slide 82 text
The near-future AI advantage (?)
My perspective on AI (and work)
How to trust and organize AI
The AI-savvy operating model
84
Slide 83
Slide 83 text
How can we empower teams of humans
+ AI agents to deliver quickly, safely, and
compliantly with high-fidelity domain
knowledge and visibility of the data
sources & results, plus the agency to
address problems caused?
85
Slide 84
Slide 84 text
End-to-end responsibility for
service outcomes is essential
to avoid harm and deliver
effectively:
use empowered teams
86
Slide 85
Slide 85 text
Start with the need for ongoing
evolution of long-lived digital
services meeting user needs…
… and work backwards from there
(don’t start with the technology)
87
Slide 86
Slide 86 text
Define guardrails and boundaries
for code gen and service specs:
domains, security, nuances of
terminology, assumptions,
algorithms, biases, etc.
(Hint: this has always been needed!)
88
Slide 87
Slide 87 text
Combine ‘architecture for flow’
(via Team Topologies) with explicit
‘active knowledge diffusion’
across flow boundaries to create
trust, alignment, and learning.
89
Slide 88
Slide 88 text
The clarity of mission and
transparency resulting from the
defined guardrails needed for
AI-augmented organizations may
be the most useful aspect of AI (!)
90
Matthew Skelton
Slide 89
Slide 89 text
Let’s do what it takes to empower teams
of humans + AI agents to make
decisions quickly, safely, and
compliantly with high-fidelity domain
knowledge and visibility of the data
sources & results, plus the agency to
address problems caused.
92
Slide 90
Slide 90 text
let’s work together
confluxhq.com
Copyright (c) 2017-2024 Conflux group of companies. All Rights Reserved.
The name “Conflux” and the filled C device are Registered Trademarks ® in multiple jurisdictions.