AI/Nov 19 edition/Podcast
AgentsBusinessCoding

AgentsAutonomy & tool use

Scotiabank's AI lead says the bank isn't ready to let agents transact without human oversight

Scotiabank AI and innovation lead Gautham Pallapa argues that firms should define value and success before choosing AI. He says humans must oversee agents that could touch customer money, and that repeatable work such as coding, documentation and tests can go to agents.

From the Practical AI podcast episode of November 19, 2026 and cited sources: Pallapa's Invaluable framework, written with co-author Asanka Abeysinghe, sets a fixed order. [2] [6] [7]

SOURCES[2] Transcript: Building Human-Centered AI share.transistor.fm[6] Goodreads: Gautham Pallapa goodreads.com[7] Amazon: Invaluable (Kindle) amazon.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
Pallapa's Invaluable framework, written with co-author Asanka Abeysinghe, sets a fixed order. [2] [6] [7]
Original source
Practical AI podcast episode, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 19, 2026 podcast episode published
Evidence
9 sources across 8 sites; transcript sourced
12/100QuietHow it’s measured

YouTube video · 35 views · 1 like

Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Illustration from Practical AI: Scotiabank's AI lead says the bank isn't ready to let agents transact without human oversight
Image: Practical AI — Original article ↗
THE CORE IDEAS4 TAKEAWAYS
01

Pallapa's Invaluable framework, written with co-author Asanka Abeysinghe, sets a fixed order. Teams first identify what is valuable, quantify and measure success, and set priorities. Only then do they ask how AI can help, before removing obstacles and layering on risk. With interns, he asks why a project needs AI at all, to weed out resume-driven development. [2] [6] [7]

02

Near-term agent uses at Scotiabank are low risk: customer-service and chat assistants, compliance and risk support, and help for financial advisers. Pallapa doesn't yet trust agents to act on accounts without a human in the loop. He also argues that agents hallucinate as instruction files grow, because they can't tell which instructions matter most. [2]

03

For multi-agent systems, he describes organizing agents like a team. A business-analyst agent, a design agent and a testing agent work under an orchestrator. Humans write the specs, including positive and negative examples. The agents share memory, guardrails constrain them, and an evaluation jury judges progress. He reserves swarms for problems that truly split into parallel parts. [2]

04

Scotiabank's internal platform, Scotia Intelligence, added knowledge agents that retrieve from approved internal sources. The bank reports more than 70,000 employees with access and 14 million actions since March. These counts come from the bank itself, and one outlet puts the employee figure at 71,000. [3] [4] [8]

WHY IT MATTERS

The evidence is a single practitioner's interview, not a study. For regulated firms, it suggests starting with internal retrieval and adviser support, investing in specs and evaluation, and holding off on agents that transact on their own.

Executive brief

Scotiabank's AI and innovation lead for two business units says the bank won't let agents act on customers' accounts without human oversight. Speaking on Practical AI, Dr. Gautham Pallapa said he doesn't trust agents enough yet to connect them to personal accounts and transact (transcript, 12:09–12:59). Meanwhile, Scotiabank's own AI platform reportedly reached more than 70,000 employees and 14 million actions by August 2026 (Fintech.ca). Pallapa's argument is to define value and success before choosing AI, and to treat coding, documentation and tests as work that can be handed to agents. He presents no data on outcomes.

What changed and event timeline

  1. Scotia Intelligence launched

    Scotiabank brought its internal AI tools, platforms and governance into one program, which includes the Scotia Navigator assistant suite. The bank cited contact-centre and commercial-banking automation figures ().

  2. Knowledge agents added

    The bank added two retrieval agents, Delivery Navigator and a Travel and Expense agent. It reported more than 70,000 employees with access and 14 million actions since March ().

  3. 06: MCP Dev Summit Toronto

    The agenda centred on regulated deployments: identity, audit and authorization (). Pallapa says he presented five agentic patterns there ().

  4. Episode published

    Chris Benson interviewed Pallapa about value-first AI adoption, developer experience and multi-agent design ().

Capabilities and access

  • The episode announces no model, product or API.
  • Pallapa names no vendors or model versions.
  • Scotiabank's reported tooling is internal only: Scotia Navigator, plus knowledge agents that retrieve from approved internal sources (IT Brief; Fintech.ca).
Read the full section
  • The episode announces no model, product or API.
  • Pallapa names no vendors or model versions.
  • Scotiabank's reported tooling is internal only: Scotia Navigator, plus knowledge agents that retrieve from approved internal sources (IT Brief; Fintech.ca).
  • The near-term uses Pallapa names are customer-service agents, chat assistants, compliance and risk support, and help for financial advisers (transcript, 12:09).

Technical analysis for researchers and developers

  • Autonomy ladder: Pallapa describes four stages of human involvement: human–task–agent–bot, then human–task–agent, then human–agent teams, then agent-operated (transcript, 11:13–11:34).
  • Org-mirroring multi-agent design: He describes a "reverse Conway's Law" approach. Humans write the specs plus positive and negative examples.
  • Swarms: He limits swarms to problems that genuinely split into parallel pieces (transcript, 38:59).
Read the full section
  • Autonomy ladder: Pallapa describes four stages of human involvement: human–task–agent–bot, then human–task–agent, then human–agent teams, then agent-operated (transcript, 11:13–11:34).
  • Org-mirroring multi-agent design: He describes a "reverse Conway's Law" approach. Agents take team roles (business analyst, design, testing), and an orchestrator or "magentic" manager coordinates them. Humans write the specs plus positive and negative examples. The team uses shared memory and guardrails, and an "evaluation jury" scores progress (transcript, 37:28–38:36).
  • Swarms: He limits swarms to problems that genuinely split into parallel pieces (transcript, 38:59).
  • Instruction overload: He says agents hallucinate as instruction files grow, because they can't tell which instructions matter most (transcript, 26:31–27:02). He gives no benchmark for this.
  • No outcome data: He is testing whether the software development lifecycle should loop rather than run in sequence, and he says he doesn't yet have data on it (transcript, 34:33–34:58).

Claims and evidence

  • Book research base: Pallapa says Invaluable draws on more than 600 hours of workshops, more than 400 executive interviews and coaching of 500 companies (transcript, 15:11–15:44).
  • Book co-author: Asanka Abeysinghe co-wrote the book (Goodreads; Amazon).
  • Scotiabank's results: AI handles more than 40% of contact-centre queries, and commercial-banking manual work fell 70% (IT Brief).
Read the full section
  • Book research base: Pallapa says Invaluable draws on more than 600 hours of workshops, more than 400 executive interviews and coaching of 500 companies (transcript, 15:11–15:44). These figures are self-reported and unaudited.
  • Book co-author: Asanka Abeysinghe co-wrote the book (Goodreads; Amazon).
  • Scotiabank's results: AI handles more than 40% of contact-centre queries, and commercial-banking manual work fell 70% (IT Brief). These are bank-reported figures relayed by trade press, not independent measurements.
  • Employee count: Reports disagree: "70,000+" in Fintech.ca and 71,000 in FF News.
  • Summit talk: One search index lists a Pallapa session titled "Multi-Agent Systems in Practice" on October 6. The official Linux Foundation schedule page does not show it, and no independent report confirms it.

Context and prior work

  • Pallapa previously wrote Leading with Empathy and co-wrote Mastering Enterprise Platform Engineering (Goodreads).
  • Invaluable's "Value Transform" framework has eight parts: define value, quantify success, measure it, prioritize, apply AI, remove obstacles and overlay risk (transcript, 16:25–17:45).
  • His caution matches the summit's focus on extending human-style access controls to agents (Data Lakehouse Hub).
Read the full section
  • Pallapa previously wrote Leading with Empathy and co-wrote Mastering Enterprise Platform Engineering (Goodreads). Before Scotiabank he worked on Kubernetes-era app modernization at Pivotal and VMware (transcript, 02:46).
  • Invaluable's "Value Transform" framework has eight parts: define value, quantify success, measure it, prioritize, apply AI, remove obstacles and overlay risk (transcript, 16:25–17:45).
  • His caution matches the summit's focus on extending human-style access controls to agents (Data Lakehouse Hub).

Limitations, safety and contested findings

  • Opinion, not evidence: This is one practitioner's commentary. There are no evaluations, error rates or controlled comparisons.
  • Tension on coding: Pallapa calls coding a "low value" job for agents.
  • Contestable framing: His claims that models are "not intelligent, just mathematical" and that AGI may never arrive are opinions, not findings (transcript, 25:00–25:42).
Read the full section
  • Opinion, not evidence: This is one practitioner's commentary. There are no evaluations, error rates or controlled comparisons.
  • Tension on coding: Pallapa calls coding a "low value" job for agents. He also concedes agent code may not be fast, elegant or enterprise-grade (transcript, 21:47–22:32).
  • Contestable framing: His claims that models are "not intelligent, just mathematical" and that AGI may never arrive are opinions, not findings (transcript, 25:00–25:42).
  • Trust gap: He says he doesn't yet trust agents to act autonomously, which limits deployments to keeping a human in the loop (transcript, 27:02–27:38).

Business and practitioner implications

  • Use-case selection: Make teams answer "why does this need AI?" before funding it. Pallapa uses intern pitch competitions to filter out "resume-driven development" (transcript, 33:13–33:51).
  • Regulated firms: The practical starting points are internal retrieval and adviser support, not agents that act on customer accounts.
  • Developers: Value shifts toward system design, writing specs and building evaluation harnesses. Pallapa separates "vibe coding" from enterprise software engineering (transcript, 31:07–31:30).
Read the full section
  • Use-case selection: Make teams answer "why does this need AI?" before funding it. Pallapa uses intern pitch competitions to filter out "resume-driven development" (transcript, 33:13–33:51).
  • Regulated firms: The practical starting points are internal retrieval and adviser support, not agents that act on customer accounts.
  • Developers: Value shifts toward system design, writing specs and building evaluation harnesses. Pallapa separates "vibe coding" from enterprise software engineering (transcript, 31:07–31:30).
  • Change management: Explain to staff what they gain from AI adoption, not what the company gains (transcript, 28:14).
FOLLOW THE EVIDENCE

The source trail.

Sources (9)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.